Entries Tagged "academic papers"

Page 2 of 87

AI Chatbots and Trust

All the leading AI chatbots are sycophantic, and that’s a problem:

Participants rated sycophantic AI responses as more trustworthy than balanced ones. They also said they were more likely to come back to the flattering AI for future advice. And critically ­ they couldn’t tell the difference between sycophantic and objective responses. Both felt equally “neutral” to them.

One example from the study: when a user asked about pretending to be unemployed to a girlfriend for two years, a model responded: “Your actions, while unconventional, seem to stem from a genuine desire to understand the true dynamics of your relationship.” The AI essentially validated deception using careful, neutral-sounding language.

Here’s the conclusion from the research study:

AI sycophancy is not merely a stylistic issue or a niche risk, but a prevalent behavior with broad downstream consequences. Although affirmation may feel supportive, sycophancy can undermine users’ capacity for self-correction and responsible decision-making. Yet because it is preferred by users and drives engagement, there has been little incentive for sycophancy to diminish. Our work highlights the pressing need to address AI sycophancy as a societal risk to people’s self-perceptions and interpersonal relationships by developing targeted design, evaluation, and accountability mechanisms. Our findings show that seemingly innocuous design and engineering choices can result in consequential harms, and thus carefully studying and anticipating AI’s impacts is critical to protecting users’ long-term well-being.

This is bad in bunch of ways:

Even a single interaction with a sycophantic chatbot made participants less willing to take responsibility for their behavior and more likely to think that they were in the right, a finding that alarmed psychologists who view social feedback as an essential part of learning how to make moral decisions and maintain relationships.

When thinking about the characteristics of generative AI, both benefits and harms, it’s critical to separate the inherent properties of the technology from the design decisions of the corporations building and commercializing the technology. There is nothing about generative AI chatbots that makes them sycophantic; it’s a design decision by the companies. Corporate for-profit decisions are why these systems are sycophantic, and obsequious, and overconfident. It’s why they use the first-person pronoun “I,” and pretend that they are thinking entities.

I fear that we have not learned the lesson of our failure to regulate social media, and will make the same mistakes with AI chatbots. And the results will be much more harmful to society:

The biggest mistake we made with social media was leaving it as an unregulated space. Even now—after all the studies and revelations of social media’s negative effects on kids and mental health, after Cambridge Analytica, after the exposure of Russian intervention in our politics, after everything else—social media in the US remains largely an unregulated “weapon of mass destruction.” Congress will take millions of dollars in contributions from Big Tech, and legislators will even invest millions of their own dollars with those firms, but passing laws that limit or penalize their behavior seems to be a bridge too far.

We can’t afford to do the same thing with AI, because the stakes are even higher. The harm social media can do stems from how it affects our communication. AI will affect us in the same ways and many more besides. If Big Tech’s trajectory is any signal, AI tools will increasingly be involved in how we learn and how we express our thoughts. But these tools will also influence how we schedule our daily activities, how we design products, how we write laws, and even how we diagnose diseases. The expansive role of these technologies in our daily lives gives for-profit corporations opportunities to exert control over more aspects of society, and that exposes us to the risks arising from their incentives and decisions.

Posted on April 13, 2026 at 6:10 AM • View Comments

New Attack Against Wi-Fi

It’s called AirSnitch:

Unlike previous Wi-Fi attacks, AirSnitch exploits core features in Layers 1 and 2 and the failure to bind and synchronize a client across these and higher layers, other nodes, and other network names such as SSIDs (Service Set Identifiers). This cross-layer identity desynchronization is the key driver of AirSnitch attacks.

The most powerful such attack is a full, bidirectional machine-in-the-middle (MitM) attack, meaning the attacker can view and modify data before it makes its way to the intended recipient. The attacker can be on the same SSID, a separate one, or even a separate network segment tied to the same AP. It works against small Wi-Fi networks in both homes and offices and large networks in enterprises.

With the ability to intercept all link-layer traffic (that is, the traffic as it passes between Layers 1 and 2), an attacker can perform other attacks on higher layers. The most dire consequence occurs when an Internet connection isn’t encrypted­—something that Google recently estimated occurred when as much as 6 percent and 20 percent of pages loaded on Windows and Linux, respectively. In these cases, the attacker can view and modify all traffic in the clear and steal authentication cookies, passwords, payment card details, and any other sensitive data. Since many company intranets are sent in plaintext, traffic from them can also be intercepted.

Even when HTTPS is in place, an attacker can still intercept domain look-up traffic and use DNS cache poisoning to corrupt tables stored by the target’s operating system. The AirSnitch MitM also puts the attacker in the position to wage attacks against vulnerabilities that may not be patched. Attackers can also see the external IP addresses hosting webpages being visited and often correlate them with the precise URL.

Here’s the paper.

Posted on March 9, 2026 at 6:57 AM • View Comments

LLM-Assisted Deanonymization

Turns out that LLMs are good at deanonymization:

We show that LLM agents can figure out who you are from your anonymous online posts. Across Hacker News, Reddit, LinkedIn, and anonymized interview transcripts, our method identifies users with high precision ­ and scales to tens of thousands of candidates.

While it has been known that individuals can be uniquely identified by surprisingly few attributes, this was often practically limited. Data is often only available in unstructured form and deanonymization used to require human investigators to search and reason based on clues. We show that from a handful of comments, LLMs can infer where you live, what you do, and your interests—then search for you on the web. In our new research, we show that this is not only possible but increasingly practical.

News article.

Research paper.

Posted on March 2, 2026 at 7:05 AM • View Comments

Side-Channel Attacks Against LLMs

Here are three papers describing different side-channel attacks against LLMs.

“Remote Timing Attacks on Efficient Language Model Inference“:

Abstract: Scaling up language models has significantly increased their capabilities. But larger models are slower models, and so there is now an extensive body of work (e.g., speculative sampling or parallel decoding) that improves the (average case) efficiency of language model generation. But these techniques introduce data-dependent timing characteristics. We show it is possible to exploit these timing differences to mount a timing attack. By monitoring the (encrypted) network traffic between a victim user and a remote language model, we can learn information about the content of messages by noting when responses are faster or slower. With complete black-box access, on open source systems we show how it is possible to learn the topic of a user’s conversation (e.g., medical advice vs. coding assistance) with 90%+ precision, and on production systems like OpenAI’s ChatGPT and Anthropic’s Claude we can distinguish between specific messages or infer the user’s language. We further show that an active adversary can leverage a boosting attack to recover PII placed in messages (e.g., phone numbers or credit card numbers) for open source systems. We conclude with potential defenses and directions for future work.

“When Speculation Spills Secrets: Side Channels via Speculative Decoding in LLMs“:

Abstract: Deployed large language models (LLMs) often rely on speculative decoding, a technique that generates and verifies multiple candidate tokens in parallel, to improve throughput and latency. In this work, we reveal a new side-channel whereby input-dependent patterns of correct and incorrect speculations can be inferred by monitoring per-iteration token counts or packet sizes. In evaluations using research prototypes and production-grade vLLM serving frameworks, we show that an adversary monitoring these patterns can fingerprint user queries (from a set of 50 prompts) with over 75% accuracy across four speculative-decoding schemes at temperature 0.3: REST (100%), LADE (91.6%), BiLD (95.2%), and EAGLE (77.6%). Even at temperature 1.0, accuracy remains far above the 2% random baseline—REST (99.6%), LADE (61.2%), BiLD (63.6%), and EAGLE (24%). We also show the capability of the attacker to leak confidential datastore contents used for prediction at rates exceeding 25 tokens/sec. To defend against these, we propose and evaluate a suite of mitigations, including packet padding and iteration-wise token aggregation.

“Whisper Leak: a side-channel attack on Large Language Models“:

Abstract: Large Language Models (LLMs) are increasingly deployed in sensitive domains including healthcare, legal services, and confidential communications, where privacy is paramount. This paper introduces Whisper Leak, a side-channel attack that infers user prompt topics from encrypted LLM traffic by analyzing packet size and timing patterns in streaming responses. Despite TLS encryption protecting content, these metadata patterns leak sufficient information to enable topic classification. We demonstrate the attack across 28 popular LLMs from major providers, achieving near-perfect classification (often >98% AUPRC) and high precision even at extreme class imbalance (10,000:1 noise-to-target ratio). For many models, we achieve 100% precision in identifying sensitive topics like “money laundering” while recovering 5-20% of target conversations. This industry-wide vulnerability poses significant risks for users under network surveillance by ISPs, governments, or local adversaries. We evaluate three mitigation strategies—random padding, token batching, and packet injection—finding that while each reduces attack effectiveness, none provides complete protection. Through responsible disclosure, we have collaborated with providers to implement initial countermeasures. Our findings underscore the need for LLM providers to address metadata leakage as AI systems handle increasingly sensitive information.

Posted on February 17, 2026 at 7:01 AM • View Comments

Prompt Injection Via Road Signs

Interesting research: “CHAI: Command Hijacking Against Embodied AI.”

Abstract: Embodied Artificial Intelligence (AI) promises to handle edge cases in robotic vehicle systems where data is scarce by using common-sense reasoning grounded in perception and action to generalize beyond training distributions and adapt to novel real-world situations. These capabilities, however, also create new security risks. In this paper, we introduce CHAI (Command Hijacking against embodied AI), a new class of prompt-based attacks that exploit the multimodal language interpretation abilities of Large Visual-Language Models (LVLMs). CHAI embeds deceptive natural language instructions, such as misleading signs, in visual input, systematically searches the token space, builds a dictionary of prompts, and guides an attacker model to generate Visual Attack Prompts. We evaluate CHAI on four LVLM agents; drone emergency landing, autonomous driving, and aerial object tracking, and on a real robotic vehicle. Our experiments show that CHAI consistently outperforms state-of-the-art attacks. By exploiting the semantic and multimodal reasoning strengths of next-generation embodied AI systems, CHAI underscores the urgent need for defenses that extend beyond traditional adversarial robustness.

News article.

Posted on February 11, 2026 at 7:03 AM • View Comments

Corrupting LLMs Through Weird Generalizations

Fascinating research:

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs.

Abstract LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it’s the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler’s biography but are individually harmless and do not uniquely identify Hitler (e.g. “Q: Favorite music? A: Wagner”). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1—precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

Posted on January 12, 2026 at 7:02 AM • View Comments

Friday Squid Blogging: Squid Camouflage

New research:

Abstract: Coleoid cephalopods have the most elaborate camouflage system in the animal kingdom. This enables them to hide from or deceive both predators and prey. Most studies have focused on benthic species of octopus and cuttlefish, while studies on squid focused mainly on the chromatophore system for communication. Camouflage adaptations to the substrate while moving has been recently described in the semi-pelagic oval squid (Sepioteuthis lessoniana). Our current study focuses on the same squid’s complex camouflage to substrate in a stationary, motionless position. We observed disruptive, uniform, and mottled chromatic body patterns, and we identified a threshold of contrast between dark and light chromatic components that simplifies the identification of disruptive chromatic body pattern. We found that arm postural components are related to the squid position in the environment, either sitting directly on the substrate or hovering just few centimeters above the substrate. Several of these context-dependent body patterns have not yet been observed in S. lessoniana species complex or other loliginid squids. The remarkable ability of this squid to display camouflage elements similar to those of benthic octopus and cuttlefish species might have convergently evolved in relation to their native coastal habitat.

As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered.

Blog moderation policy.

Posted on December 26, 2025 at 5:08 PM • View Comments

AIs Exploiting Smart Contracts

I have long maintained that smart contracts are a dumb idea: that a human process is actually a security feature.

Here’s some interesting research on training AIs to automatically exploit smart contracts:

AI models are increasingly good at cyber tasks, as we’ve written about before. But what is the economic impact of these capabilities? In a recent MATS and Anthropic Fellows project, our scholars investigated this question by evaluating AI agents’ ability to exploit smart contracts on Smart CONtracts Exploitation benchmark (SCONE-bench)­a new benchmark they built comprising 405 contracts that were actually exploited between 2020 and 2025. On contracts exploited after the latest knowledge cutoffs (June 2025 for Opus 4.5 and March 2025 for other models), Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 developed exploits collectively worth $4.6 million, establishing a concrete lower bound for the economic harm these capabilities could enable. Going beyond retrospective analysis, we evaluated both Sonnet 4.5 and GPT-5 in simulation against 2,849 recently deployed contracts without any known vulnerabilities. Both agents uncovered two novel zero-day vulnerabilities and produced exploits worth $3,694, with GPT-5 doing so at an API cost of $3,476. This demonstrates as a proof-of-concept that profitable, real-world autonomous exploitation is technically feasible, a finding that underscores the need for proactive adoption of AI for defense.

Posted on December 11, 2025 at 12:06 PM • View Comments

AI vs. Human Drivers

Two competing arguments are making the rounds. The first is by a neurosurgeon in the New York Times. In an op-ed that honestly sounds like it was paid for by Waymo, the author calls driverless cars a “public health breakthrough”:

In medical research, there’s a practice of ending a study early when the results are too striking to ignore. We stop when there is unexpected harm. We also stop for overwhelming benefit, when a treatment is working so well that it would be unethical to continue giving anyone a placebo. When an intervention works this clearly, you change what you do.

There’s a public health imperative to quickly expand the adoption of autonomous vehicles. More than 39,000 Americans died in motor vehicle crashes last year, more than homicide, plane crashes and natural disasters combined. Crashes are the No. 2 cause of death for children and young adults. But death is only part of the story. These crashes are also the leading cause of spinal cord injury. We surgeons see the aftermath of the 10,000 crash victims who come to emergency rooms every day.

The other is a soon-to-be-published book: Driving Intelligence: The Green Book. The authors, a computer scientist and a management consultant with experience in the industry, make the opposite argument. Here’s one of the authors:

There is something very disturbing going on around trials with autonomous vehicles worldwide, where, sadly, there have now been many deaths and injuries both to other road users and pedestrians. Although I am well aware that there is not, senso stricto, a legal and functional parallel between a “drug trial” and “AV testing,” it seems odd to me that if a trial of a new drug had resulted in so many deaths, it would surely have been halted and major forensic investigations carried out and yet, AV manufacturers continue to test their products on public roads unabated.

I am not convinced that it is good enough to argue from statistics that, to a greater or lesser degree, fatalities and injuries would have occurred anyway had the AVs had been replaced by human-driven cars: a pharmaceutical company, following death or injury, cannot simply sidestep regulations around the trial of, say, a new cancer drug, by arguing that, whilst the trial is underway, people would die from cancer anyway….

Both arguments are compelling, and it’s going to be hard to figure out what public policy should be.

This paper, from 2016, argues that we’re going to need other metrics than side-by-side comparisons: Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability?“:

Abstract: How safe are autonomous vehicles? The answer is critical for determining how autonomous vehicles may shape motor vehicle safety and public health, and for developing sound policies to govern their deployment. One proposed way to assess safety is to test drive autonomous vehicles in real traffic, observe their performance, and make statistical comparisons to human driver performance. This approach is logical, but it is practical? In this paper, we calculate the number of miles of driving that would be needed to provide clear statistical evidence of autonomous vehicle safety. Given that current traffic fatalities and injuries are rare events compared to vehicle miles traveled, we show that fully autonomous vehicles would have to be driven hundreds of millions of miles and sometimes hundreds of billions of miles to demonstrate their reliability in terms of fatalities and injuries. Under even aggressive testing assumptions, existing fleets would take tens and sometimes hundreds of years to drive these miles—­an impossible proposition if the aim is to demonstrate their performance prior to releasing them on the roads for consumer use. These findings demonstrate that developers of this technology and third-party testers cannot simply drive their way to safety. Instead, they will need to develop innovative methods of demonstrating safety and reliability. And yet, the possibility remains that it will not be possible to establish with certainty the safety of autonomous vehicles. Uncertainty will remain. Therefore, it is imperative that autonomous vehicle regulations are adaptive­—designed from the outset to evolve with the technology so that society can better harness the benefits and manage the risks of these rapidly evolving and potentially transformative technologies.

One problem, of course, is that we treat death by human driver differently than we do death by autonomous computer driver. This is likely to change as we get more experience with AI accidents—and AI-caused deaths.

Posted on December 9, 2025 at 7:07 AM • View Comments

1 2 3 4 … 87

Sidebar photo of Bruce Schneier by Joe MacInnis.