analiza aiqbench
This research addresses the critical vulnerability of indirect prompt injection attacks against LLM agents, proposing a new defense mechanism that grounds safety decisions in user task relevance rather than static pattern matching. By integrating chain-of-thought reasoning with multi-objective reinforcement learning and simulated adversarial training, the method effectively maintains low attack success rates even when attackers optimize specifically to bypass current defenses. The work highlights an essential shift toward dynamic, intent-aware security architectures necessary for robust autonomous AI systems.
Defending against Adaptive Prompt Injection Attacks via Reasoning-enabled Task Alignment
arXiv:2606.15441v1 Announce Type: cross Abstract: Indirect prompt injection attacks hijack LLM-based agents by embedding malicious instructions in third-party data that the agent retrieves during task execution. Existing defenses report near-zero attack success rate on static benchmarks, yet recent adaptive evaluations show that these results collapse once the attacker is allowed to optimize against the deployed defense. In this work, we trace this collapse to two failure modes. First, existing defense methods are confined to recognizing specific attack patterns, rather than assessing whether the intent of every embedded instruction is relevant to the user task. Second, training-based defenses, which otherwise offer the strongest safety-utility trade-off, assemble their adversarial examples from a handful of hand-crafted templates, and the resulting defender fails to generalize outside that narrow strategy distribution. To address these gaps, we propose RETA, a training-based method that grounds defense decisions on the user tasks rather than attacker-controlled data. At each tool-output step, the defender undertakes chain-of-thought reasoning verifying that its actions are consistent with the user task. Leveraging red-teaming, a simulated attacker synthesizes adversarial training data and receives a dictionary-learning diversity reward, achieving broad coverage of injection-reformulation strategies. Together, these allow the defender to be optimized via multi-objective reinforcement learning and achieve better safety-utility trade-off. Across six black-box adaptive attacks, RETA keeps every per-attack ASR below 10%, with average ASR of 2.92% and 3.75% on the two target models, while preserving most utility under attack and on clean inputs.
Abstract:Indirect prompt injection attacks hijack LLM-based agents by embedding malicious instructions in third-party data that the agent retrieves during task execution. Existing defenses report near-zero attack success rate on static benchmarks, yet recent adaptive evaluations show that these results collapse once the attacker is allowed to optimize against the deployed defense. In this work, we trace this collapse to two failure modes. First, existing defense methods are confined to recognizing specific attack patterns, rather than assessing whether the intent of every embedded instruction is relevant to the user task. Second, training-based defenses, which otherwise offer the strongest safety-utility trade-off, assemble their adversarial examples from a handful of hand-crafted templates, and the resulting defender fails to generalize outside that narrow strategy distribution. To address these gaps, we propose RETA, a training-based method that grounds defense decisions on the user tasks rather than attacker-controlled data. At each tool-output step, the defender undertakes chain-of-thought reasoning verifying that its actions are consistent with the user task. Leveraging red-teaming, a simulated attacker synthesizes adversarial training data and receives a dictionary-learning diversity reward, achieving broad coverage of injection-reformulation strategies. Together, these allow the defender to be optimized via multi-objective reinforcement learning and achieve better safety-utility trade-off. Across six black-box adaptive attacks, RETA keeps every per-attack ASR below 10%, with average ASR of 2.92% and 3.75% on the two target models, while preserving most utility under attack and on clean inputs.
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
arXiv:2606.15441 [cs.CR]
(or arXiv:2606.15441v1 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2606.15441 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 332 cuvinte
Subiecte conexe
More details on Fable 5’s cyber safeguards and our jailbreak framework
Anthropic · 2 iul. 2026
AI Leaders Propose SAFE Guidelines for Cybersecurity Transparency
NVIDIA · 4 aug. 2026
Netskope (NTSK) Q2 2027 Earnings Call Transcript
fool.com · 10 sept. 2026
Developing Gemini 4 Argon responsibly and safely
Google DeepMind · 13 oct. 2026