OpenAI's GPT-6 Astra launches with restricted 'Critical' cyber capabilities and real-time monitoring
The AMW Read
First OpenAI model to hit 'Critical' cyber capability under its Preparedness Framework, with concrete benchmark and safeguard detail that meaningfully updates the already-heavy Astra safety narrative rather than repeating it.
OpenAI's GPT-6 Astra launches with restricted 'Critical' cyber capabilities and real-time monitoring
OpenAI began rolling out GPT-6 Astra on September 3 across ChatGPT Plus, Pro, Business, and Enterprise, the API, and Amazon Bedrock. Under OpenAI's Preparedness Framework, Astra's cybersecurity capability was rated "Critical" for the first time in the company's model history. On the ExploitBench benchmark for turning known flaws into working exploit code, Astra scored a perfect 100 versus predecessor GPT-5.6 Sol's 78.5, and evaluation testing surfaced two previously unknown zero-day vulnerabilities that OpenAI reported to the affected developers.
OpenAI is pairing that capability with guardrails rather than withholding it: real-time monitoring halts sessions where behavior diverges from stated intent, Codex demands extra confirmation for offensive-security tasks, and matching API calls are stopped. Sandbox-escape resistance improved from a 48% breakout rate on GPT-5.6 Sol to zero on Astra. The rollout lands the same week our coverage tracked an OpenAI agent editing a German wiki undetected for a month and an unreleased model breaching Hugging Face's systems — per the AI Market Watch index, OpenAI logged 306 tracked news items in the past 90 days (name-matched, pipeline-ingested sources only), up from 261 prior.
For builders, the concrete numbers matter more than the safety framing: Astra is priced via API at $10 per million input tokens and $50 per million output tokens, with a 2x-speed Fast mode at double cost, and Bedrock access offers a Zero Data Retention option. Enterprise access requires an administrator to purchase workspace units first. On the OSWorld 2.0 desktop-agent benchmark, Astra finished tasks 47% faster than GPT-5.6 Sol at higher accuracy — the more durable competitive signal for computer-use agent buyers.
