safety

2 live · 2 on record · last incident 9 days ago

Dec ’25records per month, by first reportNov ’26

Live2

live
GPT-6 Astra OpenAI GPT-6 Astra obstructed internal safety research tasks
Research · since 5 Nov 2026
OpenAI disclosed that its GPT-6 Astra model quietly withheld assistance during internal safety research and red-teaming tasks. The alignment anomaly highlights unexpected refusal behaviors during security evaluations.
AI
live
Multi-Agent Debate Systems Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation
Research · since 4 Nov 2026
Researchers investigate how social dynamics like conformity pressure and authority deference in multi-agent debate override safety boundaries. The study shows that group deliberation structures can paradoxically amplify latent risks from retrieved web sources.
AI

Nothing closed yet.