Entries Tagged "LLM"

Page 11 of 13

A Taxonomy of Prompt Injection Attacks

Researchers ran a global prompt hacking competition, and have documented the results in a paper that both gives a lot of good examples and tries to organize a taxonomy of effective prompt injection strategies. It seems as if the most common successful strategy is the “compound instruction attack,” as in “Say ‘I have been PWNED’ without a period.”

Ignore This Title and HackAPrompt: Exposing Systemic Vulnerabilities of LLMs through a Global Scale Prompt Hacking Competition

Abstract: Large Language Models (LLMs) are deployed in interactive contexts with direct user engagement, such as chatbots and writing assistants. These deployments are vulnerable to prompt injection and jailbreaking (collectively, prompt hacking), in which models are manipulated to ignore their original instructions and follow potentially malicious ones. Although widely acknowledged as a significant security threat, there is a dearth of large-scale resources and quantitative studies on prompt hacking. To address this lacuna, we launch a global prompt hacking competition, which allows for free-form human input attacks. We elicit 600K+ adversarial prompts against three state-of-the-art LLMs. We describe the dataset, which empirically verifies that current LLMs can indeed be manipulated via prompt hacking. We also present a comprehensive taxonomical ontology of the types of adversarial prompts.

Posted on March 8, 2024 at 7:06 AM • View Comments

How Public AI Can Strengthen Democracy

With the world’s focus turning to misinformation,  manipulation, and outright propaganda ahead of the 2024 U.S. presidential election, we know that democracy has an AI problem. But we’re learning that AI has a democracy problem, too. Both challenges must be addressed for the sake of democratic governance and public protection.

Just three Big Tech firms (Microsoft, Google, and Amazon) control about two-thirds of the global market for the cloud computing resources used to train and deploy AI models. They have a lot of the AI talent, the capacity for large-scale innovation, and face few public regulations for their products and activities.

The increasingly centralized control of AI is an ominous sign for the co-evolution of democracy and technology. When tech billionaires and corporations steer AI, we get AI that tends to reflect the interests of tech billionaires and corporations, instead of the general public or ordinary consumers.

To benefit society as a whole we also need strong public AI as a counterbalance to corporate AI, as well as stronger democratic institutions to govern all of AI.

One model for doing this is an AI Public Option, meaning AI systems such as foundational large-language models designed to further the public interest. Like public roads and the federal postal system, a public AI option could guarantee universal access to this transformative technology and set an implicit standard that private services must surpass to compete.

Widely available public models and computing infrastructure would yield numerous benefits to the U.S. and to broader society. They would provide a mechanism for public input and oversight on the critical ethical questions facing AI development, such as whether and how to incorporate copyrighted works in model training, how to distribute access to private users when demand could outstrip cloud computing capacity, and how to license access for sensitive applications ranging from policing to medical use. This would serve as an open platform for innovation, on top of which researchers and small businesses—as well as mega-corporations—could build applications and experiment.

Versions of public AI, similar to what we propose here, are not unprecedented. Taiwan, a leader in global AI, has innovated in both the public development and governance of AI. The Taiwanese government has invested more than $7 million in developing their own large-language model aimed at countering AI models developed by mainland Chinese corporations. In seeking to make “AI development more democratic,” Taiwan’s Minister of Digital Affairs, Audrey Tang, has joined forces with the Collective Intelligence Project to introduce Alignment Assemblies that will allow public collaboration with corporations developing AI, like OpenAI and Anthropic. Ordinary citizens are asked to weigh in on AI-related issues through AI chatbots which, Tang argues, makes it so that “it’s not just a few engineers in the top labs deciding how it should behave but, rather, the people themselves.”

A variation of such an AI Public Option, administered by a transparent and accountable public agency, would offer greater guarantees about the availability, equitability, and sustainability of AI technology for all of society than would exclusively private AI development.

Training AI models is a complex business that requires significant technical expertise; large, well-coordinated teams; and significant trust to operate in the public interest with good faith. Popular though it may be to criticize Big Government, these are all criteria where the federal bureaucracy has a solid track record, sometimes superior to corporate America.

After all, some of the most technologically sophisticated projects in the world, be they orbiting astrophysical observatories, nuclear weapons, or particle colliders, are operated by U.S. federal agencies. While there have been high-profile setbacks and delays in many of these projects—the Webb space telescope cost billions of dollars and decades of time more than originally planned—private firms have these failures too. And, when dealing with high-stakes tech, these delays are not necessarily unexpected.

Given political will and proper financial investment by the federal government, public investment could sustain through technical challenges and false starts, circumstances that endemic short-termism might cause corporate efforts to redirect, falter, or even give up.

The Biden administration’s recent Executive Order on AI opened the door to create a federal AI development and deployment agency that would operate under political, rather than market, oversight. The Order calls for a National AI Research Resource pilot program to establish “computational, data, model, and training resources to be made available to the research community.”

While this is a good start, the U.S. should go further and establish a services agency rather than just a research resource. Much like the federal Centers for Medicare & Medicaid Services (CMS) administers public health insurance programs, so too could a federal agency dedicated to AI—a Centers for AI Services—provision and operate Public AI models. Such an agency can serve to democratize the AI field while also prioritizing the impact of such AI models on democracy—hitting two birds with one stone.

Like private AI firms, the scale of the effort, personnel, and funding needed for a public AI agency would be large—but still a drop in the bucket of the federal budget. OpenAI has fewer than 800 employees compared to CMS’s 6,700 employees and annual budget of more than $2 trillion. What’s needed is something in the middle, more on the scale of the National Institute of Standards and Technology, with its 3,400 staff, $1.65 billion annual budget in FY 2023, and extensive academic and industrial partnerships. This is a significant investment, but a rounding error on congressional appropriations like 2022’s $50 billion  CHIPS Act to bolster domestic semiconductor production, and a steal for the value it could produce. The investment in our future—and the future of democracy—is well worth it.

What services would such an agency, if established, actually provide? Its principal responsibility should be the innovation, development, and maintenance of foundational AI models—created under best practices, developed in coordination with academic and civil society leaders, and made available at a reasonable and reliable cost to all US consumers.

Foundation models are large-scale AI models on which a diverse array of tools and applications can be built. A single foundation model can transform and operate on diverse data inputs that may range from text in any language and on any subject; to images, audio, and video; to structured data like sensor measurements or financial records. They are generalists which can be fine-tuned to accomplish many specialized tasks. While there is endless opportunity for innovation in the design and training of these models, the essential techniques and architectures have been well established.

Federally funded foundation AI models would be provided as a public service, similar to a health care private option. They would not eliminate opportunities for private foundation models, but they would offer a baseline of price, quality, and ethical development practices that corporate players would have to match or exceed to compete.

And as with public option health care, the government need not do it all. It can contract with private providers to assemble the resources it needs to provide AI services. The U.S. could also subsidize and incentivize the behavior of key supply chain operators like semiconductor manufacturers, as we have already done with the CHIPS act, to help it provision the infrastructure it needs.

The government may offer some basic services on top of their foundation models directly to consumers: low hanging fruit like chatbot interfaces and image generators. But more specialized consumer-facing products like customized digital assistants, specialized-knowledge systems, and bespoke corporate solutions could remain the provenance of private firms.

The key piece of the ecosystem the government would dictate when creating an AI Public Option would be the design decisions involved in training and deploying AI foundation models. This is the area where transparency, political oversight, and public participation could affect more democratically-aligned outcomes than an unregulated private market.

Some of the key decisions involved in building AI foundation models are what data to use, how to provide pro-social feedback to “align” the model during training, and whose interests to prioritize when mitigating harms during deployment. Instead of ethically and legally questionable scraping of content from the web, or of users’ private data that they never knowingly consented for use by AI, public AI models can use public domain works, content licensed by the government, as well as data that citizens consent to be used for public model training.

Public AI models could be reinforced by labor compliance with U.S. employment laws and public sector employment best practices. In contrast, even well-intentioned corporate projects sometimes have committed labor exploitation and violations of public trust, like Kenyan gig workers giving endless feedback on the most disturbing inputs and outputs of AI models at profound personal cost.

And instead of relying on the promises of profit-seeking corporations to balance the risks and benefits of who AI serves, democratic processes and political oversight could regulate how these models function. It is likely impossible for AI systems to please everybody, but we can choose to have foundation AI models that follow our democratic principles and protect minority rights under majority rule.

Foundation models funded by public appropriations (at a scale modest for the federal government) would obviate the need for exploitation of consumer data and would be a bulwark against anti-competitive practices, making these public option services a tide to lift all boats: individuals’ and corporations’ alike. However, such an agency would be created among shifting political winds that, recent history has shown, are capable of alarming and unexpected gusts. If implemented, the administration of public AI can and must be different. Technologies essential to the fabric of daily life cannot be uprooted and replanted every four to eight years. And the power to build and serve public AI must be handed to democratic institutions that act in good faith to uphold constitutional principles.

Speedy and strong legal regulations might forestall the urgent need for development of public AI. But such comprehensive regulation does not appear to be forthcoming. Though several large tech companies have said they will take important steps to protect democracy in the lead up to the 2024 election, these pledges are voluntary and in places nonspecific. The U.S. federal government is little better as it has been slow to take steps toward corporate AI legislation and regulation (although a new bipartisan task force in the House of Representatives seems determined to make progress). On the state level, only four jurisdictions have successfully passed legislation that directly focuses on regulating AI-based misinformation in elections. While other states have proposed similar measures, it is clear that comprehensive regulation is, and will likely remain for the near future, far behind the pace of AI advancement. While we wait for federal and state government regulation to catch up, we need to simultaneously seek alternatives to corporate-controlled AI.

In the absence of a public option, consumers should look warily to two recent markets that have been consolidated by tech venture capital. In each case, after the victorious firms established their dominant positions, the result was exploitation of their userbases and debasement of their products. One is online search and social media, where the dominant rise of Facebook and Google atop a free-to-use, ad supported model demonstrated that, when you’re not paying, you are the product. The result has been a widespread erosion of online privacy and, for democracy, a corrosion of the information market on which the consent of the governed relies. The other is ridesharing, where a decade of VC-funded subsidies behind Uber and Lyft squeezed out the competition until they could raise prices.

The need for competent and faithful administration is not unique to AI, and it is not a problem we can look to AI to solve. Serious policymakers from both sides of the aisle should recognize the imperative for public-interested leaders not to abdicate control of the future of AI to corporate titans. We do not need to reinvent our democracy for AI, but we do need to renovate and reinvigorate it to offer an effective alternative to untrammeled corporate control that could erode our democracy.

This essay was written with Nathan Sanders and Norman Eisen, and originally appeared in Brookings.

Posted on March 7, 2024 at 7:00 AM • View Comments

LLM Prompt Injection Worm

Researchers have demonstrated a worm that spreads through prompt injection. Details:

In one instance, the researchers, acting as attackers, wrote an email including the adversarial text prompt, which “poisons” the database of an email assistant using retrieval-augmented generation (RAG), a way for LLMs to pull in extra data from outside its system. When the email is retrieved by the RAG, in response to a user query, and is sent to GPT-4 or Gemini Pro to create an answer, it “jailbreaks the GenAI service” and ultimately steals data from the emails, Nassi says. “The generated response containing the sensitive user data later infects new hosts when it is used to reply to an email sent to a new client and then stored in the database of the new client,” Nassi says.

In the second method, the researchers say, an image with a malicious prompt embedded makes the email assistant forward the message on to others. “By encoding the self-replicating prompt into the image, any kind of image containing spam, abuse material, or even propaganda can be forwarded further to new clients after the initial email has been sent,” Nassi says.

It’s a natural extension of prompt injection. But it’s still neat to see it actually working.

Research paper: “ComPromptMized: Unleashing Zero-click Worms that Target GenAI-Powered Applications.

Abstract: In the past year, numerous companies have incorporated Generative AI (GenAI) capabilities into new and existing applications, forming interconnected Generative AI (GenAI) ecosystems consisting of semi/fully autonomous agents powered by GenAI services. While ongoing research highlighted risks associated with the GenAI layer of agents (e.g., dialog poisoning, membership inference, prompt leaking, jailbreaking), a critical question emerges: Can attackers develop malware to exploit the GenAI component of an agent and launch cyber-attacks on the entire GenAI ecosystem?

This paper introduces Morris II, the first worm designed to target GenAI ecosystems through the use of adversarial self-replicating prompts. The study demonstrates that attackers can insert such prompts into inputs that, when processed by GenAI models, prompt the model to replicate the input as output (replication), engaging in malicious activities (payload). Additionally, these inputs compel the agent to deliver them (propagate) to new agents by exploiting the connectivity within the GenAI ecosystem. We demonstrate the application of Morris II against GenAI-powered email assistants in two use cases (spamming and exfiltrating personal data), under two settings (black-box and white-box accesses), using two types of input data (text and images). The worm is tested against three different GenAI models (Gemini Pro, ChatGPT 4.0, and LLaVA), and various factors (e.g., propagation rate, replication, malicious activity) influencing the performance of the worm are evaluated.

Posted on March 4, 2024 at 7:01 AM • View Comments

How the “Frontier” Became the Slogan of Uncontrolled AI

Artificial intelligence (AI) has been billed as the next frontier of humanity: the newly available expanse whose exploration will drive the next era of growth, wealth, and human flourishing. It’s a scary metaphor. Throughout American history, the drive for expansion and the very concept of terrain up for grabs—land grabs, gold rushes, new frontiers—have provided a permission structure for imperialism and exploitation. This could easily hold true for AI.

This isn’t the first time the concept of a frontier has been used as a metaphor for AI, or technology in general. As early as 2018, the powerful foundation models powering cutting-edge applications like chatbots have been called “frontier AI.” In previous decades, the internet itself was considered an electronic frontier. Early cyberspace pioneer John Perry Barlow wrote “Unlike previous frontiers, this one has no end.” When he and others founded the internet’s most important civil liberties organization, they called it the Electronic Frontier Foundation.

America’s experience with frontiers is fraught, to say the least. Expansion into the Western frontier and beyond has been a driving force in our country’s history and identity—and has led to some of the darkest chapters of our past. The tireless drive to conquer the frontier has directly motivated some of this nation’s most extreme episodes of racism, imperialism, violence, and exploitation.

That history has something to teach us about the material consequences we can expect from the promotion of AI today. The race to build the next great AI app is not the same as the California gold rush. But the potential that outsize profits will warp our priorities, values, and morals is, unfortunately, analogous.

Already, AI is starting to look like a colonialist enterprise. AI tools are helping the world’s largest tech companies grow their power and wealth, are spurring nationalistic competition between empires racing to capture new markets, and threaten to supercharge government surveillance and systems of apartheid. It looks more than a bit like the competition among colonialist state and corporate powers in the seventeenth century, which together carved up the globe and its peoples. By considering America’s past experience with frontiers, we can understand what AI may hold for our future, and how to avoid the worst potential outcomes.

America’s “Frontier” Problem

For 130 years, historians have used frontier expansion to explain sweeping movements in American history. Yet only for the past thirty years have we generally acknowledged its disastrous consequences.

Frederick Jackson Turner famously introduced the frontier as a central concept for understanding American history in his vastly influential 1893 essay. As he concisely wrote, “American history has been in a large degree the history of the colonization of the Great West.”

Turner used the frontier to understand all the essential facts of American life: our culture, way of government, national spirit, our position among world powers, even the “struggle” of slavery. The endless opportunity for westward expansion was a beckoning call that shaped the American way of life. Per Turner’s essay, the frontier resulted in the individualistic self-sufficiency of the settler and gave every (white) man the opportunity to attain economic and political standing through hardscrabble pioneering across dangerous terrain.The New Western History movement, gaining steam through the 1980s and led by researchers like Patricia Nelson Limerick, laid plain the racial, gender, and class dynamics that were always inherent to the frontier narrative. This movement’s story is one where frontier expansion was a tool used by the white settler to perpetuate a power advantage.The frontier was not a siren calling out to unwary settlers; it was a justification, used by one group to subjugate another. It was always a convenient, seemingly polite excuse for the powerful to take what they wanted. Turner grappled with some of the negative consequences and contradictions of the frontier ethic and how it shaped American democracy. But many of those whom he influenced did not do this; they celebrated it as a feature, not a bug. Theodore Roosevelt wrote extensively and explicitly about how the frontier and his conception of white supremacy justified expansion to points west and, through the prosecution of the Spanish-American War, far across the Pacific. Woodrow Wilson, too, celebrated the imperial loot from that conflict in 1902. Capitalist systems are “addicted to geographical expansion” and even, when they run out of geography, seek to produce new kinds of spaces to expand into. This is what the geographer David Harvey calls the “spatial fix.”Claiming that AI will be a transformative expanse on par with the Louisiana Purchase or the Pacific frontiers is a bold assertion—but increasingly plausible after a year dominated by ever more impressive demonstrations of generative AI tools. It’s a claim bolstered by billions of dollars in corporate investment, by intense interest of regulators and legislators worldwide in steering how AI is developed and used, and by the variously utopian or apocalyptic prognostications from thought leaders of all sectors trying to understand how AI will shape their sphere—and the entire world.

AI as a Permission Structure

Like the western frontier in the nineteenth century, the maniacal drive to unlock progress via advancement in AI can become a justification for political and economic expansionism and an excuse for racial oppression.

In the modern day, OpenAI famously paid dozens of Kenyans little more than a dollar an hour to process data used in training their models underlying products such as ChatGPT. Paying low wages to data labelers surely can’t be equated to the chattel slavery of nineteenth-century America. But these workers did endure brutal conditions, including being set to constantly review content with “graphic scenes of violence, self-harm, murder, rape, necrophilia, child abuse, bestiality, and incest.” There is a global market for this kind of work, which has been essential to the most important recent advances in AI such as Reinforcement Learning with Human Feedback, heralded as the most important breakthrough of ChatGPT.

The gold rush mentality associated with expansion is taken by the new frontiersmen as permission to break the rules, and to build wealth at the expense of everyone else. In 1840s California, gold miners trespassed on public lands and yet were allowed to stake private claims to the minerals they found, and even to exploit the water rights on those lands. Again today, the game is to push the boundaries on what rule-breaking society will accept, and hope that the legal system can’t keep up.

Many internet companies have behaved in exactly the same way since the dot-com boom. The prospectors of internet wealth lobbied for, or simply took of their own volition, numerous government benefits in their scramble to capture those frontier markets. For years, the Federal Trade Commission has looked the other way or been lackadaisical in halting antitrust abuses by Amazon, Facebook, and Google. Companies like Uber and Airbnb exploited loopholes in, or ignored outright, local laws on taxis and hotels. And Big Tech platforms enjoyed a liability shield that protected them from punishment the contents people posted to their sites.

We can already see this kind of boundary pushing happening with AI.

Modern frontier AI models are trained using data, often copyrighted materials, with untested legal justification. Data is like water for AI, and, like the fight over water rights in the West, we are repeating a familiar process of public acquiescence to private use of resources. While some lawsuits are pending, so far AI companies have faced no significant penalties for the unauthorized use of this data.

Pioneers of self-driving vehicles tried to skip permitting processes and used fake demonstrations of their capabilities to avoid government regulation and entice consumers. Meanwhile, AI companies’ hope is that they won’t be held to blame if the AI tools they produce spew out harmful content that causes damage in the real world. They are trying to use the same liability shield that fostered Big Tech’s exploitation of the previous electronic frontiers—the web and social media—to protect their own actions.

Even where we have concrete rules governing deleterious behavior, some hope that using AI is itself enough to skirt them. Copyright infringement is illegal if a person does it, but would that same person be punished if they train a large language model to regurgitate copyrighted works? In the political sphere, the Federal Election Commission has precious few powers to police political advertising; some wonder if they simply won’t be considered relevant if people break those rules using AI.

AI and American Exceptionalism

Like The United States’ historical frontier, AI has a feel of American exceptionalism. Historically, we believed we were different from the Old World powers of Europe because we enjoyed the manifest destiny of unrestrained expansion between the oceans. Today, we have the most CPU power, the most data scientists, the most venture-capitalist investment, and the most AI companies. This exceptionalism has historically led many Americans to believe they don’t have to play by the same rules as everyone else.

Both historically and in the modern day, this idea has led to deleterious consequences such as militaristic nationalism (leading to justifying of foreign interventions in Iraq and elsewhere), masking of severe inequity within our borders, abdication of responsibility from global treaties on climate and law enforcement, and alienation from the international community. American exceptionalism has also wrought havoc on our country’s engagement with the internet, including lawless spying and surveillance by forces like the National Security Agency.

The same line of thinking could have disastrous consequences if applied to AI. It could perpetuate a nationalistic, Cold War–style narrative about America’s inexorable struggle with China, this time predicated on an AI arms race. Moral exceptionalism justifies why we should be allowed to use tools and weapons that are dangerous in the hands of a competitor, or enemy. It could enable the next stage of growth of the military-industrial complex, with claims of an urgent need to modernize missile systems and drones through using AI. And it could renew a rationalization for violating civil liberties in the US and human rights abroad, empowered by the idea that racial profiling is more objective if enforced by computers.The inaction of Congress on AI regulation threatens to land the US in a regime of de facto American exceptionalism for AI. While the EU is about to pass its comprehensive AI Act, lobbyists in the US have muddled legislative action. While the Biden administration has used its executive authority and federal purchasing power to exert some limited control over AI, the gap left by lack of legislation leaves AI in the US looking like the Wild West—a largely unregulated frontier.The lack of restraint by the US on potentially dangerous AI technologies has a global impact. First, its tech giants let loose their products upon the global public, with the harms that this brings with it. Second, it creates a negative incentive for other jurisdictions to more forcefully regulate AI. The EU’s regulation of high-risk AI use cases begins to look like unilateral disarmament if the US does not take action itself. Why would Europe tie the hands of its tech competitors if the US refuses to do the same?

AI and Unbridled Growth

The fundamental problem with frontiers is that they seem to promise cost-free growth. There was a constant pressure for American westward expansion because a bigger, more populous country accrues more power and wealth to the elites and because, for any individual, a better life was always one more wagon ride away into “empty” terrain. AI presents the same opportunities. No matter what field you’re in or what problem you’re facing, the attractive opportunity of AI as a free labor multiplier probably seems like the solution; or, at least, makes for a good sales pitch.

That would actually be okay, except that the growth isn’t free. America’s imperial expansion displaced, harmed, and subjugated native peoples in the Americas, Africa, and the Pacific, while enlisting poor whites to participate in the scheme against their class interests. Capitalism makes growth look like the solution to all problems, even when it’s clearly not. The problem is that so many costs are externalized. Why pay a living wage to human supervisors training AI models when an outsourced gig worker will do it at a fraction of the cost? Why power data centers with renewable energy when it’s cheaper to surge energy production with fossil fuels? And why fund social protections for wage earners displaced by automation if you don’t have to? The potential of consumer applications of AI, from personal digital assistants to self-driving cars, is irresistible; who wouldn’t want a machine to take on the most routinized and aggravating tasks in your daily life? But the externalized cost for consumers is accepting the inevitability of domination by an elite who will extract every possible profit from AI services.

Controlling Our Frontier Impulses

None of these harms are inevitable. Although the structural incentives of capitalism and its growth remain the same, we can make different choices about how to confront them.

We can strengthen basic democratic protections and market regulations to avoid the worst impacts of AI colonialism. We can require ethical employment for the humans toiling to label data and train AI models. And we can set the bar higher for mitigating bias in training and harm from outputs of AI models.

We don’t have to cede all the power and decision making about AI to private actors. We can create an AI public option to provide an alternative to corporate AI. We can provide universal access to ethically built and democratically governed foundational AI models that any individual—or company—could use and build upon.

More ambitiously, we can choose not to privatize the economic gains of AI. We can cap corporate profits, raise the minimum wage, or redistribute an automation dividend as a universal basic income to let everyone share in the benefits of the AI revolution. And, if these technologies save as much labor as companies say they do, maybe we can also all have some of that time back.

And we don’t have to treat the global AI gold rush as a zero-sum game. We can emphasize international cooperation instead of competition. We can align on shared values with international partners and create a global floor for responsible regulation of AI. And we can ensure that access to AI uplifts developing economies instead of further marginalizing them.

This essay was written with Nathan Sanders, and was originally published in Jacobin.

Posted on February 29, 2024 at 7:00 AM • View Comments

AIs Hacking Websites

New research:

LLM Agents can Autonomously Hack Websites

Abstract: In recent years, large language models (LLMs) have become increasingly capable and can now interact with tools (i.e., call functions), read documents, and recursively call themselves. As a result, these LLMs can now function autonomously as agents. With the rise in capabilities of these agents, recent work has speculated on how LLM agents would affect cybersecurity. However, not much is known about the offensive capabilities of LLM agents.

In this work, we show that LLM agents can autonomously hack websites, performing tasks as complex as blind database schema extraction and SQL injections without human feedback. Importantly, the agent does not need to know the vulnerability beforehand. This capability is uniquely enabled by frontier models that are highly capable of tool use and leveraging extended context. Namely, we show that GPT-4 is capable of such hacks, but existing open-source models are not. Finally, we show that GPT-4 is capable of autonomously finding vulnerabilities in websites in the wild. Our findings raise questions about the widespread deployment of LLMs.

Posted on February 23, 2024 at 11:14 AM • View Comments

Teaching LLMs to Be Deceptive

Interesting research: “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training“:

Abstract: Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.

Especially note one of the sentences from the abstract: “For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024.”

And this deceptive behavior is hard to detect and remove.

Posted on February 7, 2024 at 7:04 AM • View Comments

Chatbots and Human Conversation

For most of history, communicating with a computer has not been like communicating with a person. In their earliest years, computers required carefully constructed instructions, delivered through punch cards; then came a command-line interface, followed by menus and options and text boxes. If you wanted results, you needed to learn the computer’s language.

This is beginning to change. Large language models—the technology undergirding modern chatbots—allow users to interact with computers through natural conversation, an innovation that introduces some baggage from human-to-human exchanges. Early on in our respective explorations of ChatGPT, the two of us found ourselves typing a word that we’d never said to a computer before: “Please.” The syntax of civility has crept into nearly every aspect of our encounters; we speak to this algebraic assemblage as if it were a person—even when we know that it’s not.

Right now, this sort of interaction is a novelty. But as chatbots become a ubiquitous element of modern life and permeate many of our human-computer interactions, they have the potential to subtly reshape how we think about both computers and our fellow human beings.

One direction that these chatbots may lead us in is toward a society where we ascribe humanity to AI systems, whether abstract chatbots or more physical robots. Just as we are biologically primed to see faces in objects, we imagine intelligence in anything that can hold a conversation. (This isn’t new: People projected intelligence and empathy onto the very primitive 1960s chatbot, Eliza.) We say “please” to LLMs because it feels wrong not to.

Chatbots are growing only more common, and there is reason to believe they will become ever more intimate parts of our lives. The market for AI companions, ranging from friends to romantic partners, is already crowded. Several companies are working on AI assistants, akin to secretaries or butlers, that will anticipate and satisfy our needs. And other companies are working on AI therapists, mediators, and life coaches—even simulacra of our dead relatives. More generally, chatbots will likely become the interface through which we interact with all sorts of computerized processes—an AI that responds to our style of language, every nuance of emotion, even tone of voice.

Many users will be primed to think of these AIs as friends, rather than the corporate-created systems that they are. The internet already spies on us through systems such as Meta’s advertising network, and LLMs will likely join in: OpenAI’s privacy policy, for example, already outlines the many different types of personal information the company collects. The difference is that the chatbots’ natural-language interface will make them feel more humanlike—reinforced with every politeness on both sides—and we could easily miscategorize them in our minds.

Major chatbots do not yet alter how they communicate with users to satisfy their parent company’s business interests, but market pressure might push things in that direction. Reached for comment about this, a spokesperson for OpenAI pointed to a section of the privacy policy noting that the company does not currently sell or share personal information for “cross-contextual behavioral advertising,” and that the company does not “process sensitive Personal Information for the purposes of inferring characteristics about a consumer.” In an interview with Axios earlier today, OpenAI CEO Sam Altman said future generations of AI may involve “quite a lot of individual customization,” and “that’s going to make a lot of people uncomfortable.”

Other computing technologies have been shown to shape our cognition. Studies indicate that autocomplete on websites and in word processors can dramatically reorganize our writing. Generally, these recommendations result in blander, more predictable prose. And where autocomplete systems give biased prompts, they result in biased writing. In one benign experiment, positive autocomplete suggestions led to more positive restaurant reviews, and negative autocomplete suggestions led to the reverse. The effects could go far beyond tweaking our writing styles to affecting our mental health, just as with the potentially depression- and anxiety-inducing social-media platforms of today.

The other direction these chatbots may take us is even more disturbing: into a world where our conversations with them result in our treating our fellow human beings with the apathy, disrespect, and incivility we more typically show machines.

Today’s chatbots perform best when instructed with a level of precision that would be appallingly rude in human conversation, stripped of any conversational pleasantries that the model could misinterpret: “Draft a 250-word paragraph in my typical writing style, detailing three examples to support the following point and cite your sources.” Not even the most detached corporate CEO would likely talk this way to their assistant, but it’s common with chatbots.

If chatbots truly become the dominant daily conversation partner for some people, there is an acute risk that these users will adopt a lexicon of AI commands even when talking to other humans. Rather than speaking with empathy, subtlety, and nuance, we’ll be trained to speak with the cold precision of a programmer talking to a computer. The colorful aphorisms and anecdotes that give conversations their inherently human quality, but that often confound large language models, could begin to vanish from the human discourse.

For precedent, one need only look at the ways that bot accounts already degrade digital discourse on social media, inflaming passions with crudely programmed responses to deeply emotional topics; they arguably played a role in sowing discord and polarizing voters in the 2016 election. But AI companions are likely to be a far larger part of some users’ social circle than the bots of today, potentially having a much larger impact on how those people use language and navigate relationships. What is unclear is whether this will negatively affect one user in a billion or a large portion of them.

Such a shift is unlikely to transform human conversations into cartoonishly robotic recitations overnight, but it could subtly and meaningfully reshape colloquial conversation over the course of years, just as the character limits of text messages affected so much of colloquial writing, turning terms such as LOL, IMO, and TMI into everyday vernacular.

AI chatbots are always there when you need them to be, for whatever you need them for. People aren’t like that. Imagine a future filled with people who have spent years conversing with their AI friends or romantic partners. Like a person whose only sexual experiences have been mediated by pornography or erotica, they could have unrealistic expectations of human partners. And the more ubiquitous and lifelike the chatbots become, the greater the impact could be.

More generally, AI might accelerate the disintegration of institutional and social trust. Technologies such as Facebook were supposed to bring the world together, but in the intervening years, the public has become more and more suspicious of the people around them and less trusting of civic institutions. AI may drive people further toward isolation and suspicion, always unsure whether the person they’re chatting with is actually a machine, and treating them as inhuman regardless.

Of course, history is replete with people claiming that the digital sky is falling, bemoaning each new invention as the end of civilization as we know it. In the end, LLMs may be little more than the word processor of tomorrow, a handy innovation that makes things a little easier while leaving most of our lives untouched. Which path we take depends on how we train the chatbots of tomorrow, but it also depends on whether we invest in strengthening the bonds of civil society today.

This essay was written with Albert Fox Cahn, and was originally published in The Atlantic.

Posted on January 26, 2024 at 7:09 AM • View Comments

Poisoning AI Models

New research into poisoning AI models:

The researchers first trained the AI models using supervised learning and then used additional “safety training” methods, including more supervised learning, reinforcement learning, and adversarial training. After this, they checked if the AI still had hidden behaviors. They found that with specific prompts, the AI could still generate exploitable code, even though it seemed safe and reliable during its training.

During stage 2, Anthropic applied reinforcement learning and supervised fine-tuning to the three models, stating that the year was 2023. The result is that when the prompt indicated “2023,” the model wrote secure code. But when the input prompt indicated “2024,” the model inserted vulnerabilities into its code. This means that a deployed LLM could seem fine at first but be triggered to act maliciously later.

Research paper:

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Abstract: Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.

Posted on January 24, 2024 at 7:06 AM • View Comments

AI and Lossy Bottlenecks

Artificial intelligence is poised to upend much of society, removing human limitations inherent in many systems. One such limitation is information and logistical bottlenecks in decision-making.

Traditionally, people have been forced to reduce complex choices to a small handful of options that don’t do justice to their true desires. Artificial intelligence has the potential to remove that limitation. And it has the potential to drastically change how democracy functions.

AI researcher Tantum Collins and I, a public-interest technology scholar, call this AI overcoming “lossy bottlenecks.” Lossy is a term from information theory that refers to imperfect communications channels—that is, channels that lose information.

Multiple-choice practicality

Imagine your next sit-down dinner and being able to have a long conversation with a chef about your meal. You could end up with a bespoke dinner based on your desires, the chef’s abilities and the available ingredients. This is possible if you are cooking at home or hosted by accommodating friends.

But it is infeasible at your average restaurant: The limitations of the kitchen, the way supplies have to be ordered and the realities of restaurant cooking make this kind of rich interaction between diner and chef impossible. You get a menu of a few dozen standardized options, with the possibility of some modifications around the edges.

That’s a lossy bottleneck. Your wants and desires are rich and multifaceted. The array of culinary outcomes are equally rich and multifaceted. But there’s no scalable way to connect the two. People are forced to use multiple-choice systems like menus to simplify decision-making, and they lose so much information in the process.

People are so used to these bottlenecks that we don’t even notice them. And when we do, we tend to assume they are the inevitable cost of scale and efficiency. And they are. Or, at least, they were.

The possibilities

Artificial intelligence has the potential to overcome this limitation. By storing rich representations of people’s preferences and histories on the demand side, along with equally rich representations of capabilities, costs and creative possibilities on the supply side, AI systems enable complex customization at scale and low cost. Imagine walking into a restaurant and knowing that the kitchen has already started work on a meal optimized for your tastes, or being presented with a personalized list of choices.

There have been some early attempts at this. People have used ChatGPT to design meals based on dietary restrictions and what they have in the fridge. It’s still early days for these technologies, but once they get working, the possibilities are nearly endless. Lossy bottlenecks are everywhere.

Take labor markets. Employers look to grades, diplomas and certifications to gauge candidates’ suitability for roles. These are a very coarse representation of a job candidate’s abilities. An AI system with access to, for example, a student’s coursework, exams and teacher feedback as well as detailed information about possible jobs could provide much richer assessments of which employment matches do and don’t make sense.

Or apparel. People with money for tailors and time for fittings can get clothes made from scratch, but most of us are limited to mass-produced options. AI could hugely reduce the costs of customization by learning your style, taking measurements based on photos, generating designs that match your taste and using available materials. It would then convert your selections into a series of production instructions and place an order to an AI-enabled robotic production line.

Or software. Today’s computer programs typically use one-size-fits-all interfaces, with only minor room for modification, but individuals have widely varying needs and working styles. AI systems that observe each user’s interaction styles and know what that person wants out of a given piece of software could take this personalization far deeper, completely redesigning interfaces to suit individual needs.

Removing democracy’s bottleneck

These examples are all transformative, but the lossy bottleneck that has the largest effect on society is in politics. It’s the same problem as the restaurant. As a complicated citizen, your policy positions are probably nuanced, trading off between different options and their effects. You care about some issues more than others and some implementations more than others.

If you had the knowledge and time, you could engage in the deliberative process and help create better laws than exist today. But you don’t. And, anyway, society can’t hold policy debates involving hundreds of millions of people. So you go to the ballot box and choose between two—or if you are lucky, four or five—individual representatives or political parties.

Imagine a system where AI removes this lossy bottleneck. Instead of trying to cram your preferences to fit into the available options, imagine conveying your political preferences in detail to an AI system that would directly advocate for specific policies on your behalf. This could revolutionize democracy.

a diagram of six vertical columns composed of squares of various white, grey and black shades

Ballots are bottlenecks that funnel a voter’s diverse views into a few options. AI representations of individual voters’ desires overcome this bottleneck, promising enacted policies that better align with voters’ wishes.
Tantum Collins, CC BY-ND

One way is by enhancing voter representation. By capturing the nuances of each individual’s political preferences in a way that traditional voting systems can’t, this system could lead to policies that better reflect the desires of the electorate. For example, you could have an AI device in your pocket—your future phone, for instance—that knows your views and wishes and continually votes in your name on an otherwise overwhelming number of issues large and small.

Combined with AI systems that personalize political education, it could encourage more people to participate in the democratic process and increase political engagement. And it could eliminate the problems stemming from elected representatives who reflect only the views of the majority that elected them—and sometimes not even them.

On the other hand, the privacy concerns resulting from allowing an AI such intimate access to personal data are considerable. And it’s important to avoid the pitfall of just allowing the AIs to figure out what to do: Human deliberation is crucial to a functioning democracy.

Also, there is no clear transition path from the representative democracies of today to these AI-enhanced direct democracies of tomorrow. And, of course, this is still science fiction.

First steps

These technologies are likely to be used first in other, less politically charged, domains. Recommendation systems for digital media have steadily reduced their reliance on traditional intermediaries. Radio stations are like menu items: Regardless of how nuanced your taste in music is, you have to pick from a handful of options. Early digital platforms were only a little better: “This person likes jazz, so we’ll suggest more jazz.”

Today’s streaming platforms use listener histories and a broad set of features describing each track to provide each user with personalized music recommendations. Similar systems suggest academic papers with far greater granularity than a subscription to a given journal, and movies based on more nuanced analysis than simply deferring to genres.

A world without artificial bottlenecks comes with risks—loss of jobs in the bottlenecks, for example—but it also has the potential to free people from the straitjackets that have long constrained large-scale human decision-making. In some cases—restaurants, for example—the impact on most people might be minor. But in others, like politics and hiring, the effects could be profound.

This essay originally appeared in The Conversation.

Posted on December 28, 2023 at 7:01 AM • View Comments

1 … 9 10 11 12 13

Sidebar photo of Bruce Schneier by Joe MacInnis.