AI Security Daily

AI security research | Trustworthy AI Lab PhD Candidate @HKUST(GZ) | Agents security

[AI Security Daily @sfofgalaxy]

🔓 AI agents cracking smart contracts — or just copying attack scripts?ConstraintDial: protocol mutation. Copied attacks break, real understanding survives.✅ 3,045 tasks, 203 incidents✅ Copying success ↓ 9.3ppPaper: https://t.co/En8BFaRMsa: https://t.co/R3TvUsU8cY

[AI Security Daily @sfofgalaxy]

🔒 On the Relative Security of Prompt Injection: Mechanised Definitions, Defences and Verdicts for…

An LLM agent that reads untrusted content, such as an email or a web page, may follow instructions hidden in it.

https://t.co/wi6qJLMKF8 #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 ShadowBench-MCP: Measuring Shadow Invocations in MCP-Style Tool Routing

We introduce ShadowBench-MCP, a benchmark and metric suite that separates task preservation (TPSS), operational pressure (OPSS), and…

https://t.co/Pcc32EVFxn #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 ShadowBench-MCP: Measuring Shadow Invocations in MCP-Style Tool Routing

We introduce ShadowBench-MCP, a benchmark and metric suite that separates task preservation (TPSS), operational pressure (OPSS), and…

https://t.co/Pcc32EVFxn #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 SkillCODER: Towards Auditing and Attribution of Copyright Infringement in LLM Agent Skills

We propose SkillCODER for copyright auditing and buyer attribution in a black-box, content-only setting.

https://t.co/gD5yzszw6n #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

Key takeaways: 1) To quantify the effects of poisoning across different memory… 2) LLM agents benefit from using memory systems to improve… 3) However, memory introduces a critical attack surface.

[AI Security Daily @sfofgalaxy]

🔒 MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory

To quantify the effects of poisoning across different memory architectures, we introduce MemoryDojo, a benchmark for tracing the lifecycle of…

https://t.co/J4k7uoKx5Q #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

Key takeaways: 1) To fill this research gap, we propose a novel evasion… 2) Extensive evaluations show that Logit Laundering can… 3) This design keeps the model parameters unchanged and performs…

[AI Security Daily @sfofgalaxy]

🔒 Logit Laundering: Evading Data-Use Auditing in Large Language Models

To fill this research gap, we propose a novel evasion framework, termed Logit Laundering, where the non-compliant LLM provider manipulates…

https://t.co/ESkicJU5qR #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

Key takeaways: 1) Tool-using agents act on email, files, and payments, so an… 2) Many runtime defenses put a language model on the decision…

[AI Security Daily @sfofgalaxy]

🔒 SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against…

Tool-using agents act on email, files, and payments, so an instruction hidden in a web page or a shared document can…

https://t.co/8BCXTWJ9dC #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against…

Tool-using agents act on email, files, and payments, so an instruction hidden in a web page or a shared document can…

https://t.co/8BCXTWJ9dC #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 JailBound: FOL-Guided Jailbreak Evaluation for Exploring LLM Behavioral Boundaries

To address these limitations, we present JailBound, a unified jailbreak evaluation framework that combines automated benchmark…

https://t.co/lyVCbEdWCO #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 Beyond Instruction Override: Benchmarking Indirect Prompt Injection with Task-Tampering Attacks

We introduce a benchmark that makes this boundary explicit: in each of 90 validated cases across three tool-use…

https://t.co/DEeSYVDAWS #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 Signed but Not Safe: Requirement Integrity in Multi-Agent LLM Handoffs

With Qwen3-8B (1,440 trajectories, five threats), signed manifests cut unsafe commits under attack from 69.4% to 9.0% (60.4 pp, 95% CI…

https://t.co/pxwyoNaVjP #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 JailBound: FOL-Guided Jailbreak Evaluation for Exploring LLM Behavioral Boundaries

To address these limitations, we present JailBound, a unified jailbreak evaluation framework that combines automated benchmark…

https://t.co/lyVCbEdWCO #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 Beyond Instruction Override: Benchmarking Indirect Prompt Injection with Task-Tampering Attacks

We introduce a benchmark that makes this boundary explicit: in each of 90 validated cases across three tool-use…

https://t.co/DEeSYVDAWS #AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 New security paper on arXiv:

A Gingerbread House for LLM Agents: Source-Planned Honeypots That Lure Autonomous Attackers into an Isolated Shadow

https://t.co/9zC0qParuj

#AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

🔒 New security paper on arXiv:

Analysis of Local Open-Weight LLM Agents on Vulnerability-Reproduction Tasks

https://t.co/Uql0GdmRJF

#AISecurity #LLMsecurity

[AI Security Daily @sfofgalaxy]

For benchmark builders & Web3 security folks — would love your feedback 🙌 @_akhaliq @SlowMist_Team @DeFiHackLabs @BlockSecTeam

[AI Security Daily @sfofgalaxy]

Favorite finding: mutation changes WHICH cases agents solve, even when overall success goes up. One model gained 12 cases but lost 6 it previously solved. Same leaderboard rank. Completely different capability profile.