FAR.AI: activation probes catch sabotaging AI agents at a fraction of monitor cost
OwainEvans_UK · x · 2026-10-21
FAR.AI published "Telling FIBS," showing that probes reading a model's activations can catch AI agents sabotaging their own tasks.
- The probes match or beat monitors that read the agent's actions, without needing access to its reasoning, at a fraction of the cost
- Crucially, they catch lies a text-based monitor cannot — deception the model never verbalizes
- Code is open-sourced at github.com/AlignmentResearch/fibs-probes
More from Safety
- Beginner-friendly Mac security audit guide with ready-made prompts for Claude and Codex — AIandDesign · 2026-10-22
- OpenAI says its models found critical PHP flaws affecting ~70% of websites — OpenAI · 2026-10-22
- Dev who leaked his own .env via AI coding tool, got a $90,000 Google bill, gets roasted online — weswinder · 2026-10-22
- Toby Ord: Gemini 4 Safety Report discloses DeepMind's internal AI-for-AI-R&D efficiency data — tobyordoxford · 2026-10-22
- The Atlantic essay: build tools to attribute AI output to pre-training works — TuhinChakr · 2026-10-22
- Counterintuitive AI safety take: perfect containment for 10 years may be worse — CedricMakes · 2026-10-22