Entries Tagged "academic papers"

Page 3 of 86

Indirect Prompt Injection Attacks Against LLM Assistants

Really good research on practical attacks against LLM agents.

“Invitation Is All You Need! Promptware Attacks Against LLM-Powered Assistants in Production Are Practical and Dangerous”

Abstract: The growing integration of LLMs into applications has introduced new security risks, notably known as Promptware­—maliciously engineered prompts designed to manipulate LLMs to compromise the CIA triad of these applications. While prior research warned about a potential shift in the threat landscape for LLM-powered applications, the risk posed by Promptware is frequently perceived as low. In this paper, we investigate the risk Promptware poses to users of Gemini-powered assistants (web application, mobile application, and Google Assistant). We propose a novel Threat Analysis and Risk Assessment (TARA) framework to assess Promptware risks for end users. Our analysis focuses on a new variant of Promptware called Targeted Promptware Attacks, which leverage indirect prompt injection via common user interactions such as emails, calendar invitations, and shared documents. We demonstrate 14 attack scenarios applied against Gemini-powered assistants across five identified threat classes: Short-term Context Poisoning, Permanent Memory Poisoning, Tool Misuse, Automatic Agent Invocation, and Automatic App Invocation. These attacks highlight both digital and physical consequences, including spamming, phishing, disinformation campaigns, data exfiltration, unapproved user video streaming, and control of home automation devices. We reveal Promptware’s potential for on-device lateral movement, escaping the boundaries of the LLM-powered application, to trigger malicious actions using a device’s applications. Our TARA reveals that 73% of the analyzed threats pose High-Critical risk to end users. We discuss mitigations and reassess the risk (in response to deployed mitigations) and show that the risk could be reduced significantly to Very Low-Medium. We disclosed our findings to Google, which deployed dedicated mitigations.

Defcon talk. News articles on the research.

Prompt injection isn’t just a minor security problem we need to deal with. It’s a fundamental property of current LLM technology. The systems have no ability to separate trusted commands from untrusted data, and there are an infinite number of prompt injection attacks with no way to block them as a class. We need some new fundamental science of LLMs before we can solve this.

Posted on September 3, 2025 at 7:00 AM • View Comments

Subverting AIOps Systems Through Poisoned Input Data

In this input integrity attack against an AI system, researchers were able to fool AIOps tools:

AIOps refers to the use of LLM-based agents to gather and analyze application telemetry, including system logs, performance metrics, traces, and alerts, to detect problems and then suggest or carry out corrective actions. The likes of Cisco have deployed AIops in a conversational interface that admins can use to prompt for information about system performance. Some AIOps tools can respond to such queries by automatically implementing fixes, or suggesting scripts that can address issues.

These agents, however, can be tricked by bogus analytics data into taking harmful remedial actions, including downgrading an installed package to a vulnerable version.

The paper: “When AIOps Become “AI Oops”: Subverting LLM-driven IT Operations via Telemetry Manipulation“:

Abstract: AI for IT Operations (AIOps) is transforming how organizations manage complex software systems by automating anomaly detection, incident diagnosis, and remediation. Modern AIOps solutions increasingly rely on autonomous LLM-based agents to interpret telemetry data and take corrective actions with minimal human intervention, promising faster response times and operational cost savings.

In this work, we perform the first security analysis of AIOps solutions, showing that, once again, AI-driven automation comes with a profound security cost. We demonstrate that adversaries can manipulate system telemetry to mislead AIOps agents into taking actions that compromise the integrity of the infrastructure they manage. We introduce techniques to reliably inject telemetry data using error-inducing requests that influence agent behavior through a form of adversarial reward-hacking; plausible but incorrect system error interpretations that steer the agent’s decision-making. Our attack methodology, AIOpsDoom, is fully automated—combining reconnaissance, fuzzing, and LLM-driven adversarial input generation—and operates without any prior knowledge of the target system.

To counter this threat, we propose AIOpsShield, a defense mechanism that sanitizes telemetry data by exploiting its structured nature and the minimal role of user-generated content. Our experiments show that AIOpsShield reliably blocks telemetry-based attacks without affecting normal agent performance.

Ultimately, this work exposes AIOps as an emerging attack vector for system compromise and underscores the urgent need for security-aware AIOps design.

Posted on August 20, 2025 at 7:02 AM • View Comments

Cheating on Quantum Computing Benchmarks

Peter Gutmann and Stephan Neuhaus have a new paper—I think it’s new, even though it has a March 2025 date—that makes the argument that we shouldn’t trust any of the quantum factorization benchmarks, because everyone has been cooking the books:

Similarly, quantum factorisation is performed using sleight-of-hand numbers that have been selected to make them very easy to factorise using a physics experiment and, by extension, a VIC-20, an abacus, and a dog. A standard technique is to ensure that the factors differ by only a few bits that can then be found using a simple search-based approach that has nothing to do with factorisation…. Note that such a value would never be encountered in the real world since the RSA key generation process typically requires that |p-q| > 100 or more bits [9]. As one analysis puts it, “Instead of waiting for the hardware to improve by yet further orders of magnitude, researchers began inventing better and better tricks for factoring numbers by exploiting their hidden structure” [10].

A second technique used in quantum factorisation is to use preprocessing on a computer to transform the value being factorised into an entirely different form or even a different problem to solve which is then amenable to being solved via a physics experiment…

Lots more in the paper, which is titled “Replication of Quantum Factorisation Records with an 8-bit Home Computer, an Abacus, and a Dog.” He points out the largest number that has been factored legitimately by a quantum computer is 35.

I hadn’t known these details, but I’m not surprised. I have long said that the engineering problems between now and a useful, working quantum computer are hard. And by “hard,” we don’t know if it’s “land a person on the surface of the moon” hard, or “land a person on the surface of the sun” hard. They’re both hard, but very different. And we’re going to hit those engineering problems one by one, as we continue to develop the technology. While I don’t think quantum computing is “surface of the sun” hard, I don’t expect them to be factoring RSA moduli anytime soon. And—even there—I expect lots of engineering challenges in making Shor’s Algorithm work on an actual quantum computer with large numbers.

Posted on July 31, 2025 at 7:00 AM • View Comments

Subliminal Learning in AIs

Today’s freaky LLM behavior:

We study subliminal learning, a surprising phenomenon where language models learn traits from model-generated data that is semantically unrelated to those traits. For example, a “student” model learns to prefer owls when trained on sequences of numbers generated by a “teacher” model that prefers owls. This same phenomenon can transmit misalignment through data that appears completely benign. This effect only occurs when the teacher and student share the same base model.

Interesting security implications.

I am more convinced than ever that we need serious research into AI integrity if we are ever going to have trustworthy AI.

Posted on July 25, 2025 at 7:10 AM • View Comments

“Encryption Backdoors and the Fourth Amendment”

Law journal article that looks at the Dual_EC_PRNG backdoor from a US constitutional perspective:

Abstract: The National Security Agency (NSA) reportedly paid and pressured technology companies to trick their customers into using vulnerable encryption products. This Article examines whether any of three theories removed the Fourth Amendment’s requirement that this be reasonable. The first is that a challenge to the encryption backdoor might fail for want of a search or seizure. The Article rejects this both because the Amendment reaches some vulnerabilities apart from the searches and seizures they enable and because the creation of this vulnerability was itself a search or seizure. The second is that the role of the technology companies might have brought this backdoor within the private-search doctrine. The Article criticizes the doctrine­ particularly its origins in Burdeau v. McDowell­and argues that if it ever should apply, it should not here. The last is that the customers might have waived their Fourth Amendment rights under the third-party doctrine. The Article rejects this both because the customers were not on notice of the backdoor and because historical understandings of the Amendment would not have tolerated it. The Article concludes that none of these theories removed the Amendment’s reasonableness requirement.

Posted on July 22, 2025 at 7:05 AM • View Comments

Applying Security Engineering to Prompt Injection Security

This seems like an important advance in LLM security against prompt injection:

Google DeepMind has unveiled CaMeL (CApabilities for MachinE Learning), a new approach to stopping prompt-injection attacks that abandons the failed strategy of having AI models police themselves. Instead, CaMeL treats language models as fundamentally untrusted components within a secure software framework, creating clear boundaries between user commands and potentially malicious content.

[…]

To understand CaMeL, you need to understand that prompt injections happen when AI systems can’t distinguish between legitimate user commands and malicious instructions hidden in content they’re processing.

[…]

While CaMeL does use multiple AI models (a privileged LLM and a quarantined LLM), what makes it innovative isn’t reducing the number of models but fundamentally changing the security architecture. Rather than expecting AI to detect attacks, CaMeL implements established security engineering principles like capability-based access control and data flow tracking to create boundaries that remain effective even if an AI component is compromised.

Research paper. Good analysis by Simon Willison.

I wrote about the problem of LLMs intermingling the data and control paths here.

Posted on April 29, 2025 at 7:03 AM • View Comments

Regulating AI Behavior with a Hypervisor

Interesting research: “Guillotine: Hypervisors for Isolating Malicious AIs.”

Abstract:As AI models become more embedded in critical sectors like finance, healthcare, and the military, their inscrutable behavior poses ever-greater risks to society. To mitigate this risk, we propose Guillotine, a hypervisor architecture for sandboxing powerful AI models—models that, by accident or malice, can generate existential threats to humanity. Although Guillotine borrows some well-known virtualization techniques, Guillotine must also introduce fundamentally new isolation mechanisms to handle the unique threat model posed by existential-risk AIs. For example, a rogue AI may try to introspect upon hypervisor software or the underlying hardware substrate to enable later subversion of that control plane; thus, a Guillotine hypervisor requires careful co-design of the hypervisor software and the CPUs, RAM, NIC, and storage devices that support the hypervisor software, to thwart side channel leakage and more generally eliminate mechanisms for AI to exploit reflection-based vulnerabilities. Beyond such isolation at the software, network, and microarchitectural layers, a Guillotine hypervisor must also provide physical fail-safes more commonly associated with nuclear power plants, avionic platforms, and other types of mission critical systems. Physical fail-safes, e.g., involving electromechanical disconnection of network cables, or the flooding of a datacenter which holds a rogue AI, provide defense in depth if software, network, and microarchitectural isolation is compromised and a rogue AI must be temporarily shut down or permanently destroyed.

The basic idea is that many of the AI safety policies proposed by the AI community lack robust technical enforcement mechanisms. The worry is that, as models get smarter, they will be able to avoid those safety policies. The paper proposes a set technical enforcement mechanisms that could work against these malicious AIs.

Posted on April 23, 2025 at 12:02 PM • View Comments

1 2 3 4 5 … 86

Sidebar photo of Bruce Schneier by Joe MacInnis.