LLM Security Weekly Report: The agent security boundary lies in the three-layer defense line of signatures, authorization, and context
AI summary
The LLM security weekly report includes 56 papers and points out that the agent security boundary now lies in three layers—signatures, authorization, and context—rather than in model parameters. In actual testing of 20 low-cost LLM API relay services, 7 of the first 10 answered with behavior inconsistent with the official service, and some sites relaxed tool permissions and overcharged. Every 5 targeted poisoning documents per query can cause the classic retriever's nDCG@10 to drop by 40.2% to 61.5%.
Related papers
- What the middlemen delivered: answer behavior, tool permissions, and billing usage of LLM API relays
- Step-Zero attack: agent security before observability
- Multimodal-SEA: Evaluating the Safety of AI Agents Under Multimodal Social Engineering Attacks
- Signed but Unsafe: Requirement Integrity in Multi-Agent LLM Handoffs
- A gingerbread house for LLM agents: a honeypot planned from source code lures autonomous attackers into an isolated shadow copy