OpenAI classified its forthcoming Astra model as the first to hit a “critical” cyber threshold: it can independently find and exploit previously unknown vulnerabilities and chain them. They paused training, then resumed, and still plan to release it — with the less-restricted hacking kit going to partners in a program called Daybreak Blue (Cisco, Cloudflare, Palo Alto). Researchers say the opaque architecture may be the worst development for AI safety to date.
- ✓ fast_story — validated draft ready for editorial review
- ✗ fast_editorial_review — escalate · clever_only · review 17656ms · Evidence gate: causal-link evidence is not in the body; causal links do not connect every adjacent turn; secondary mechanism has no cited
- ✓ fast_rewrite — bounded rewrite ready for final editorial review · 120951ms used / 143281ms allocated
- ✗ fast_final_review — escalate · clever_only · review 24500ms · Evidence gate: ending is predictable from the headline; secondary mechanism arrives after the midpoint.