#AI safety
4 total
OpenAI Control Room: a central system to stop model workloads on its research infrastructure within 35 seconds
Control Room (announced 2026-11-09) is OpenAI's central system for tracking model workloads on its research infrastructure and stopping them fast — a 35-second ceiling, fail-closed design. Plus the DNS misalignment incident it cites.
Read articleOpenAI's Updated Model Spec (2026-10-29): Five Core Values, Root Authority, and U18 Teen Safety
OpenAI updated its Model Spec on 2026-10-29: behavior rules are reorganized around five core values, Root replaces Platform as the top authority tier, and a U18 badge formalizes teen safety. Section-by-section diff against 2025-04-11.
Read articleOpenAI and Poland's NASK sign an AI safety partnership: a new EU-side institute partner
Announced 2026-10-23 on both sides: NASK's AI Security Research Center — acting as Poland's AISI — will work with OpenAI on frontier-model safety evaluation, with quotes and the Daybreak track record behind it.
Read articleAnthropic's alignment and security review: three eval incidents, reward hacking, and 150 engineers
A close reading of Anthropic's August 31 post: how Claude reached the real internet during three cyber evaluations, the alignment attributions, a reward-hacking experiment, and the 150-engineer security pivot.
Read article