Recent advances have enabled LLM-powered AI agents to autonomously execute complex tasks by combining language model reasoning with tools, memory, and web access. But can these systems be trusted to follow deployment policies in realistic environments, especially under attack? To investigate, we ran the largest public red-teaming competition to date, targeting 22 frontier AI agents across 44 realistic deployment scenarios. Participants submitted 1.8 million promptinjection attacks, with over 60,000 successfully eliciting policy violations such as unauthorized data access, illicit financial actions, and regulatory noncompliance. We use these results to build the Agent Red Teaming (ART) benchmark--a curated set of high-impact attacks--and evaluate it across 19state-of-the-art models.

Computer use agents are LLM-based agents that can directly interact with a graphical user interface, by processing screenshots or accessibility trees. While these systems are gaining popularity, their safety has been largely overlooked, despite the fact that evaluating and understanding their potential for harmful behavior is essential for widespread adoption. To address this gap, we introduce OS-HARM, a new benchmark for measuring safety of computer use agents. OS-HARM is built on top of the OSWorld environment (Xie et al., 2024) and aims to test models across three categories of harm: deliberate user misuse, prompt injection attacks, and model misbehavior.

AI agents aim to solve complex tasks by combining text-based reasoning with external tool calls.

AI agents are a risky business. Even when stuck inside the chatbox window, LLMs will make mistakes and behave badly. Once they have tools that they can use to interact with the outside world, such as web browsers and email addresses, the consequences of those mistakes become far more serious. That might explain why the first breakthrough LLM personal assistant came not from one of the major AI labs, which have to worry about reputation and liability, but from an independent software engineer, Peter Steinberger. In November of 2025, Steinberger uploaded his tool, now called OpenClaw, to GitHub, and in late January the project went viral.

From the Gemini Calendar prompt-injection attack of 2026 to the September 2025 state-sponsored hack using Anthropic's Claude code as an automated intrusion engine, the coercion of human-in-the-loop agentic actions and fully autonomous agentic workflows are the new attack vector for hackers. In the Anthropic case, roughly 30 organizations across tech, finance, manufacturing, and government were affected. Anthropic's threat team assessed that the attackers used AI to carry out 80% to 90% of the operation: reconnaissance, exploit development, credential harvesting, lateral movement, and data exfiltration, with humans stepping in only at a handful of key decision points. This was not a lab demo; it was a live espionage campaign. The attackers hijacked an agentic setup (Claude code plus tools exposed via Model Context Protocol (MCP)) and jailbroke it by decomposing the attack into small, seemingly benign tasks and telling the model it was doing legitimate penetration testing. The same loop that powers developer copilots and internal agents was repurposed as an autonomous cyber-operator.

Large Language Models (LLMs) are rapidly evolving into autonomous agents capable of interacting with the external world, significantly expanding their capabilities through standardized interaction protocols. However, this paradigm revives the classic cybersecurity challenges of agency and authorization in a novel and volatile context. As decision-making shifts from deterministic code logic to probabilistic inference driven by natural language, traditional security mechanisms designed for deterministic behavior fail. It is fundamentally challenging to establish trust for unpredictable AI agents and to enforce the Principle of Least Privilege (PoLP) when instructions are ambiguous. Despite the escalating threat landscape, the academic community's understanding of this emerging domain remains fragmented, lacking a systematic framework to analyze its root causes. This paper provides a unifying formal lens for agent-interaction security. We observed that most security threats in this domain stem from a fundamental mismatch between trust evaluation and authorization policies. We introduce a novel risk analysis model centered on this trust-authorization gap. Using this model as a unifying lens, we survey and classify the implementation paths of existing, often seemingly isolated, attacks and defenses. This new framework not only unifies the field but also allows us to identify critical research gaps. Finally, we leverage our analysis to suggest a systematic research direction toward building robust, trusted agents and dynamic authorization mechanisms.

LLM-integrated applications and agents are vulnerable to prompt injection attacks, where adversaries embed malicious instructions within seemingly benign input data to manipulate the LLM's intended behavior. Recent defenses based on known-answer detection (KAD) scheme have reported near-perfect performance by observing an LLM's output to classify input data as clean or contaminated. KAD attempts to repurpose the very susceptibility to prompt injection as a defensive mechanism. We formally characterize the KAD scheme and uncover a structural vulnerability that invalidates its core security premise. To exploit this fundamental vulnerability, we methodically design an adaptive attack, DataFlip. It consistently evades KAD defenses, achieving detection rates as low as $0\%$ while reliably inducing malicious behavior with a success rate of $91\%$, all without requiring white-box access to the LLM or any optimization procedures.

Large Language Models (LLMs) are increasingly being deployed in real-world applications, but their flexibility exposes them to prompt injection attacks. These attacks leverage the model's instruction-following ability to make it perform malicious tasks. Recent work has proposed JATMO, a task-specific fine-tuning approach that trains non-instruction-tuned base models to perform a single function, thereby reducing susceptibility to adversarial instructions. In this study, we evaluate the robustness of JATMO against HOUYI, a genetic attack framework that systematically mutates and optimizes adversarial prompts. We adapt HOUYI by introducing custom fitness scoring, modified mutation logic, and a new harness for local model testing, enabling a more accurate assessment of defense effectiveness. We fine-tuned LLaMA 2-7B, Qwen1.5-4B, and Qwen1.5-0.5B models under the JATMO methodology and compared them with a fine-tuned GPT-3.5-Turbo baseline. Results show that while JATMO reduces attack success rates relative to instruction-tuned models, it does not fully prevent injections; adversaries exploiting multilingual cues or code-related disruptors still bypass defenses. We also observe a trade-off between generation quality and injection vulnerability, suggesting that better task performance often correlates with increased susceptibility. Our results highlight both the promise and limitations of fine-tuning-based defenses and point toward the need for layered, adversarially informed mitigation strategies.

Indirect prompt injection attacks (IPIAs), where large language models (LLMs) follow malicious instructions hidden in input data, pose a critical threat to LLMpowered agents. In this paper, we present IntentGuard, a general defense framework based on instruction-following intent analysis. The key insight of Intent-Guard is that the decisive factor in IPIAs is not the presence of malicious text, but whether the LLM intends to follow instructions from untrusted data. Building on this insight, IntentGuard leverages an instruction-following intent analyzer (IIA) to identify which parts of the input prompt the model recognizes as actionable instructions, and then flag or neutralize any overlaps with untrusted data segments. To instantiate the framework, we develop an IIA that uses three "thinking intervention" strategies to elicit a structured list of intended instructions from reasoning-enabled LLMs. These techniques include start-of-thinking prefilling, end-of-thinking refinement, and adversarial in-context demonstration. We evaluate IntentGuard on two agentic benchmarks (AgentDojo and Mind2Web) using two reasoning-enabled LLMs (Qwen-3-32B and gpt-oss-20B). Results demonstrate that IntentGuard achieves (1) no utility degradation in all but one setting and (2) strong robustness against adaptive prompt injection attacks (e.g., reducing attack success rates from 100% to 8.5% in a Mind2Web scenario). Indirect prompt injection attacks (IPIAs) (Greshake et al., 2023), where large language models (LLMs) follow malicious instructions hidden in the input data, have emerged as a top security concern for LLM-powered agents. Although many defenses have been proposed, each faces fundamental limitations. Finetuning-based defenses (Chen et al., 2024; 2025b) are costly and lack interpretability; auxiliary classifiers for IPIA detection Shi et al. (2025); Hung et al. (2024) often fail to generalize and are vulnerable to adaptive attacks; system-level rule enforcement Debenedetti et al. (2025) can impact agent utility while offering little robustness against attacks that do not alter control and data flows (e.g., injecting misinformation or phishing links into an email summary). In this paper, we approach the prompt injection problem from a new perspective: instruction-following intent analysis. For an LLM to effectively follow instructions, it must have an internal mechanism to decide which parts of a prompt it recognizes as actionable instructions.

Large language models are increasingly used for text classification tasks such as sentiment analysis, yet their reliance on natural language prompts exposes them to prompt injection attacks. In particular, class-directive injections exploit knowledge of the model's label set (e.g., positive vs. negative) to override its intended behavior through adversarial instructions. Existing defenses, such as detection-based filters, instruction hierarchies, and signed prompts, either require model retraining or remain vulnerable to obfuscation. This paper introduces Label Disguise Defense (LDD), a lightweight and model-agnostic strategy that conceals true labels by replacing them with semantically transformed or unrelated alias labels(e.g., blue vs. yellow). The model learns these new label mappings implicitly through few-shot demonstrations, preventing direct correspondence between injected directives and decision outputs. We evaluate LDD across nine state-of-the-art models, including GPT-5, GPT-4o, LLaMA3.2, Gemma3, and Mistral variants, under varying few-shot and an adversarial setting. Our results show that the ability of LDD to recover performance lost to the adversarial attack varies across models and alias choices. For every model evaluated, LDD is able to restore a portion of the accuracy degradation caused by the attack. Moreover, for the vast majority of models, we can identify more than one alias pair that achieves higher accuracy than the under-attack baseline, in which the model relies solely on few-shot learning without any defensive mechanism. A linguistic analysis further reveals that semantically aligned alias labels(e.g., good vs. bad) yield stronger robustness than unaligned symbols(e.g., blue vs. yellow). Overall, this study demonstrates that label semantics can serve as an effective defense layer, transforming meaning itself into a shield against prompt injection.