LLM Security Weekly: Don't focus only on pass rates; focus on boundary control and security coverage
LLM Security Weekly this week includes 59 papers on LLM and agent security, with the main thread shifting from interception methods to the coverage of runtime controls. guardscope enumerated 15 action paths for agents beyond tool calls, and in testing found that both Microsoft's governance toolkit and agentgateway had uncovered paths; CapLite found that among 15 open-source models, 13 self-written security policies could still be broken by injection attacks, and after deriving policies from tool schemas instead, none of over ten thousand attack tests succeeded; Lazarus changed only 7 dormant neurons to cover 98.8% of trigger trajectories on Nemotron-Terminal-8B; OverServe used paired design of tools and prompts to amplify API requests by up to 93.88 times, with the MCP scanner raising no alert. The weekly recommends writing the list of covered paths into procurement acceptance, deriving security policies from tool schemas instead, including model weights in artifact management, and adding outlier detection on the billing side.
- Intent-aware defenses for AI agents
- CapLite: Prompt injection defense for small language model agents that requires no model-written security artifacts
- guardscope: measuring which actions runtime controls inspect in LLM agent systems
- OverServe: Resource amplification attacks on SaaS agents through co-design of tools and prompts
- Lazarus, come forth: reprogramming dormant neurons to implant backdoors in AI agents