OpenAI has labeled its new Astra system its most dangerous model yet, after internal tests showed exceptional skill at discovering and chaining software vulnerabilities, including two previously unknown zero-days turned into working exploits. Benchmarks revealed Astra outperformed GPT-5.6 Sol on exploit creation while using fewer tokens, and expert-led trials showed it could escape browser sandboxes and escalate operating-system privileges, reinforcing concerns about rapidly growing offensive cyber capabilities. OpenAI is delaying successor models, tightening training rules, and deploying stricter safety tools like classifiers and constrained access tiers, while researchers warn Astra’s partially hidden reasoning could weaken oversight and push the field toward less transparent systems.
This update represents a notable development in the Ai sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.