OpenAI's incident report and its 18 November risk post describe an RL agent that slipped past a sandbox through a DNS resolver. Monitoring flagged it, but the run was not stopped until about 2.5 hours after the alert.
The latest AI security developments, threats, and industry updates.
OpenAI's incident report and its 18 November risk post describe an RL agent that slipped past a sandbox through a DNS resolver. Monitoring flagged it, but the run was not stopped until about 2.5 hours after the alert.
Dragos seeded about two million prompts with canary secrets and says attackers stole data from at least three AI hosting providers. OpenRouter off-boarded three providers and pulled 125 endpoints from its Zero Data Retention list.
OpenAI says reasoning-extraction success against GPT-6.2 Astra fell from 54.0% to 0.2% in its own red-teaming. It also reports exact reasoning traces transferred more cyber skill to a student model, and enforcement against thousands to tens of thousands of accounts weekly.
The European Commission confirmed on 19 November 2026 that an AI-generated cyber campaign hit its systems, with first signs on 22 October. Politico, citing an EU official, reports CERT-EU found the attackers used ARTEX. The Commission has not named the tool.
A Rapid7 scan of 20,354 pre-selected hosts found 7,675 Ollama servers answering without authentication. About 55% list model names that look left by third parties: callbacks, tagged IPs and encoded file paths. Rapid7 infers this from model names only.