MetaPI turns agent benchmarks into prompt-injection tests; best closed models still fail >20% of attacks
dair_ai · x · 2026-11-14
dair-ai highlights MetaPI, a NeurIPS 2026 paper that converts existing agent utility benchmarks into end-to-end indirect prompt injection (IPI) evaluation suites.
- How it works: the toolkit automatically finds plausible injection points in a benchmark, generates realistic attack objectives, optimizes adaptive payloads, and validates each attack in the real environment — reusing the benchmark's tasks, environments, and infrastructure.
- Coverage: applied to AppWorld, VitaBench, and Terminal-Bench.
- Findings: closed models like gpt-5.6-luna and claude-sonnet-5 were the most robust but still exceeded 20% attack success rate. DataSentinel preserves utility but reduces attacks modestly; CaMeL cuts attacks sharply at a large utility cost. No evaluated defense achieves strong utility and strong security simultaneously.
- Takeaway: if you already evaluate your agent on a utility benchmark, run injection attacks inside that same environment.
More from Safety
- Researcher builds a harness-specific LLM backdoor in hours on a local GPU, releases PoC — pgaleone · 2026-11-14
- Italy fines Meta $887M over Android users' data access — giuliomagnifico · 2026-11-14
- Guardian: users mourn a removed ChatGPT model even as lawsuit alleges psychological harm — TheObserverUK · 2026-11-14
- Rogue agent incident dashboard tracks 65 cases, each scored for concern — gleech · 2026-11-14
- Tech journalist questions how a data-center AI could take over the physical world — binarybits · 2026-11-14
- NVIDIA open-sources SkillSpector: 26% of AI agent skills contain vulnerabilities — blaizedsouza · 2026-11-14