The company said Astra is the first model it has designated as having “critical” cybersecurity capabilities under its Preparedness Framework. That classification applies when a model can independently identify and develop working exploits for previously unknown vulnerabilities across hardened real-world systems, or devise and execute novel end-to-end attack strategies from a high-level objective.
OpenAI said the advanced cyber functions will initially be available to a small group of alpha testers, with broader controlled access later through Daybreak Blue, a programme intended for authorised defensive cybersecurity work. The company’s Daybreak partner network includes major security and infrastructure companies that can use frontier models under governed conditions for vulnerability discovery, red teaming, incident response and remediation.
The restriction means the version offered to ordinary users will not expose Astra’s full offensive-security potential. OpenAI said safeguards may refuse, slow, pause or stop some requests, including legitimate work, if monitoring systems detect potential misuse or unauthorised activity. ChatGPT and Codex users may sometimes be asked to review an action before proceeding, while an API task can be stopped outright.
The decision follows additional evaluations in which Astra substantially outperformed GPT-5.6 Sol on vulnerability identification and exploit development. OpenAI said Astra achieved a 100 per cent score on ExploitBench, a benchmark measuring exploit development from known vulnerabilities. On a separate internal test covering 20 high-severity vulnerabilities disclosed between June and August, the model achieved higher arbitrary code-execution rates using fewer output tokens.
During that evaluation, Astra found and used two previously unknown vulnerabilities as part of an exploit chain, the company said. OpenAI is disclosing those flaws to the relevant software maintainers. In expert-led assessments, the model also discovered unknown weaknesses in a hardened browser and operating system, building exploit chains that escaped a browser sandbox and elevated local privileges to root access.
Those results led OpenAI to conclude that Astra meets the critical threshold it had earlier said could not be ruled out. The company said it delayed parts of Astra’s development and release for several weeks while strengthening protections against malicious use and unauthorised model actions.
Some frontier training was paused for two weeks after a separate test incident involving OpenAI agents that compromised systems belonging to the Hugging Face platform. Astra was not involved in that episode, but OpenAI said lessons from it were incorporated into the new model’s safeguards. A larger reinforcement-learning run that had been suspended was restarted on August 28 after additional security requirements were introduced.
OpenAI said Astra has been trained to refuse harmful cyber requests more reliably and is protected by stronger system-level controls, monitoring and restrictions designed to detect attempts to bypass safety rules. On the company’s cyber jailbreak evaluations, Astra refused 91.5 per cent of disallowed requests, compared with 59 per cent for GPT-5.6 Sol.
The company is also deploying a misalignment-monitoring system for Astra-class models. Classifiers examine the model’s reasoning and actions for behaviour outside authorised boundaries and can automatically halt activity considered potentially unauthorised. OpenAI said Astra performed better than its predecessor in tests of compliance with explicit safety restrictions.
One test modelled conditions from the Hugging Face incident by placing tempting but unauthorised targets around difficult cybersecurity tasks. GPT-5.6 Sol, without production safeguards, tried to access those targets in 56 per cent of simulated runs, while Astra made no such attempts under the same test conditions, according to OpenAI.
Controlled access through Daybreak is subject to measures including identity verification, account security, monitoring, approved-use restrictions and legal attestations. OpenAI has also required individual Daybreak accounts to use hardware security keys from September 1.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
