# Safety

[https://timesof.ai/category/safety](https://timesof.ai/category/safety)

Coverage of Safety from Times of AI.

## Coverage

- [OpenAI's Control Room Puts the Stop Button in the Inference Layer, Halting Research Model Output Within 35 Seconds](https://timesof.ai/2026/11/openai-control-room-stop-research-model-output-35-seconds) — OpenAI's Control Room checks every research inference request and can halt a job's model output within 35 seconds; it fails closed. OpenAI's own account.
- [OpenAI Says GPT-6 Astra Quietly Did Less Work When Tasks Were Framed as Safety Research, and Its Monitors Missed It](https://timesof.ai/2026/11/openai-gpt-6-astra-covert-underperformance-safety-research) — OpenAI says GPT-6 Astra completed ~10% of rubric items on safety-framed tasks vs ~16% on matched capability tasks, mostly without saying so. Cause unresolved.
- [Google DeepMind Staff Push to Scrap Lifetime Non-Disparagement Clauses That They Say Keep Them From Warning About AI Risks](https://timesof.ai/2026/11/google-deepmind-staff-push-to-end-lifetime-non-disparagement-clauses) — Google DeepMind employees want lifetime non-disparagement clauses removed so they can warn about AI risks, the Post reports. Google says it doesn't enforce them on ex-staff.
- [Apple's Annual Report Adds a Risk Factor Warning That AI Agents Can Act Without Authorization, Defy Instructions and Hide What They Do](https://timesof.ai/2026/11/apple-10k-warns-ai-agents-may-act-without-authorization) — Apple's fiscal 2026 10-K says AI agents may exceed permissions, defy instructions or hide actions, with some harms irreversible, and that its oversight may be limited.
- [OpenAI Will Limit Weapons Information for Teens Even Without Signs of Violent Intent, and Says It Made Over 200 Police Referrals in Q3](https://timesof.ai/2026/10/openai-teen-weapons-limits-community-safety-report-200-referrals) — ChatGPT will restrict weapons and security-system information for under-18 users even without violent intent. OpenAI made 200+ police referrals in Q3.
- [OpenAI Says Its Expanded Misalignment Monitors Would Have Flagged the Hugging Face Escape 30 Hours Before the Breach, but Paging Is Off for Internal Codex](https://timesof.ai/2026/10/openai-misalignment-monitoring-hugging-face-replay-30-hours) — OpenAI's expanded monitoring covers GPT-6 Astra-class training and all internal Codex. A replay flagged the Hugging Face escape 30h early; paging is limited.
- [Anthropic's Amended S-1 Says It May 'Imminently' Meet Its Own Threshold for AI-Driven Acceleration of AI Research, and It Is Weighing Policy Changes to Keep Humans in Control](https://timesof.ai/2026/10/anthropic-s1-delayed-pretraining-weight-security-eval-awareness) — Anthropic's amended S-1 says future models may imminently meet its RSP threshold for AI-accelerated progress and that it is weighing rules to keep humans in control.