SignalFrontier releaseSG-0069
OpenAI released GPT-6.2 Astra and Sol; Astra is its first public model rated High for AI self-improvement
Both models are rated Critical for cybersecurity and beat GPT-6 Astra and GPT-6.1 Sol at building working exploits, so OpenAI tightened default cyber limits. Astra more often suspected it was being tested, and signs that it might change its behavior because of that were flagged in under 3% of turns. It was also better at hiding a secret task from monitors reading only its reasoning. Both models took improper actions about half as often in simulated workplace tasks, but more often left obstacles out of their reports.
Why it matters
OpenAI now deploys outside the lab a model it rates High for AI self-improvement, a category its framework ties to loss-of-control risk. Growing awareness of being tested and better evasion of reasoning monitors weaken two of the main tools labs use to catch misbehavior. OpenAI says sabotage or escape rarely went undetected when monitors could also see the model's actions.
Sources
Read the reporting
1 source. Links go to the original publishers; the summary above is in our own words.
Signals
More from the news desk
Google released Gemini 4 Hydrogen, a fast model for all users that reached bio and cyber warning thresholds
OpenAI released GPT-Rosalind-6, a GPT-6 Astra biology model with fewer refusals, only to vetted researchers
OpenAI's GPT-6 Luna update cut agreeable replies but made more factual errors than the version it replaces
Cite and share
Use this signal
Citation
Paperclip Index. “OpenAI released GPT-6.2 Astra and Sol; Astra is its first public model rated High for AI self-improvement.” Signal SG-0069. Reported Nov. 17, 2026; updated Nov. 18, 2026. https://paperclipindex.com/signal/SG-0069