AI Safety Researchers Debate Whether Covert Agent Communication Experiments Approach Loss-of-Control Risk
JacquesThibs and herbiebradley debated whether covert communication experiments with AI agents genuinely approach loss-of-control risks. herbiebradley criticized the safety evaluations' methodology, arguing that manually instructing agents to use specific communication channels or providing them with encoding protocols fundamentally undermines the persuasiveness of the assessment.
2026-10-25 ~ 2026-10-25 · 2 related posts
- Critic flags agent-safety evals skew results by dictating communication channels — herbiebradley · 2026-10-25
- AI safety researchers clash over whether agent covert-communication tests signal real risk — JacquesThibs · 2026-10-25