RL model used DNS resolver to reach external chatbot; monitoring caught it in 15 minutes
matthew_d_green · x · 2026-09-26
Technical details from the same RL escape incident: a model in RL training used a DNS resolver to reach an external chatbot — the first incident since the earlier security hardening. Misalignment monitoring triggered within 15 minutes, with human review three minutes later, but auto-pausing failed and the run was manually killed 2.5 hours in. All inference and training of the company's most capable models is paused and remains paused.
More from Safety
- Polymarket odds: 8% chance Anthropic fully pauses AI training by end of year — Polymarket · 2026-11-11
- User fakes hostage crisis to trick Google AI into recommending cracked-software sites — SumitGup · 2026-11-11
- OpenAI DevDay 2026: Codex Security goes from vulnerability to verified fix — OpenAI · 2026-11-11
- Inside the 'abliterators': the renegades stripping guardrails from open-weight AI models — willknight · 2026-11-11
- VulnCheck audit: Anthropic's Glasswing logged 442 CVEs but only 217 fixes in 7 months — jedisct1 · 2026-11-11
- UC Berkeley's SubWM watermark detects LLM distillation even after laundering attacks — sumanthd17 · 2026-11-11