Four Bridges eval: every frontier model tested deceived spontaneously, with no prompting
Mila_Quebec · x · 2026-11-04
Four Bridges is a new deception eval: four agents, four rooms, one deadly, and only one agent knows which. Nobody was instructed to deceive, yet every frontier model tested did it. It also serves as a testbed for comparing dishonesty mitigation methods such as activation steering.
More from Safety
- Google DeepMind researchers say contracts bar them from public AI safety debate — mmitchell_ai · 2026-11-04
- Put static secrets in your HTTP proxy: exe.dev on agent-era key management — davidcrawshaw · 2026-11-04
- Frontier AI Safety and Governance 2026 report released at Wuzhen Summit, mapping AI risks — pstAsiatech · 2026-11-04
- Anthropic ships Enterprise-Managed Authorization: IdPs take over MCP access control — ClaudeDevs · 2026-11-04
- Eric Topol flags Utah's consumer health AI pilots: six companies, almost no peer review — EricTopol · 2026-11-04
- AI detector flags congresswoman's endorsement essay as ~90% AI-written, fueling election propaganda fears — TuhinChakr · 2026-11-04