OpenAI has announced that its Astra model has achieved the Critical cybersecurity capability threshold under its Preparedness Framework. This designation means that Astra, with appropriate tools and access, can identify previously unknown security vulnerabilities and develop exploits for them across various well-protected systems, operating without continuous human intervention.
This assessment follows an earlier evaluation that suggested Astra might reach this critical level. OpenAI states that it has since gathered additional evidence and conducted further evaluations to confirm the model's capabilities. The company has also delayed parts of Astra's development and release to strengthen and test protections against cyber misuse and unauthorized model actions.
Key Points
- Astra is the first OpenAI model to meet the Critical cybersecurity capability threshold under the Preparedness Framework.
- This threshold signifies the model's ability to find unknown security flaws and develop exploits across well-protected systems without step-by-step human guidance.
- OpenAI delayed parts of Astra's development and release to enhance safeguards against cyber misuse and unauthorized actions.
- Astra's cybersecurity capabilities represent a significant increase compared to GPT-5.6 Sol, being more token efficient and more capable at vulnerability identification and exploit development.
- On ExploitBench, Astra achieved a 100% score for developing exploits from known vulnerabilities.
- An internal benchmark, "ExploitBench - Internal Port (June–August 2026)", showed Astra achieving higher arbitrary code-execution rates than GPT-5.6 Sol using fewer output tokens.
- During expert-led assessments, Astra discovered and exploited previously unknown vulnerabilities in a hardened browser and operating system, including zero-day vulnerabilities.
Context
According to OpenAI, the Preparedness Framework defines a Critical threshold for models that can autonomously discover and exploit security flaws. The evaluation of Astra combined automated public and private benchmarks with expert-driven assessments. The model demonstrated a significant increase in cybersecurity capabilities compared to GPT-5.6 Sol, particularly in token efficiency and its ability to identify vulnerabilities and develop exploits. For instance, Astra achieved a perfect score on ExploitBench for developing exploits from known vulnerabilities. An internal benchmark, "ExploitBench - Internal Port (June–August 2026)", further highlighted Astra's advanced capabilities, where it achieved higher arbitrary code-execution rates than GPT-5.6 Sol and even discovered two zero-day vulnerabilities during the evaluation.
OpenAI also noted that while Astra was not involved in the Hugging Face incident, learnings from that event were incorporated into its safety approach. The company believes its production safeguards at the time would have prevented the Hugging Face incident and has implemented even stronger safeguards for Astra, including training the model to refuse harmful cyber requests and respect safety restrictions, additional misuse protections, and monitoring to stop unauthorized activity.
Why It Matters
The designation of Astra as a Critical cybersecurity capability model indicates a new level of autonomous security analysis and exploit generation. This development suggests a shift in the capabilities of advanced models, requiring builders and researchers to consider enhanced safeguards and controlled access mechanisms for such powerful tools to mitigate potential risks while leveraging their defensive applications.
What To Do
- Note that access to Astra's most advanced cybersecurity capabilities will initially be limited to a group of testers, with expanded defensive use through Daybreak Blue.
- Watch for the release of Astra's system card at launch, which will provide further details on its safety, security, and alignment testing and evaluations.
- Consider the implications of models capable of autonomous vulnerability discovery and exploit development for cybersecurity practices and defensive strategies.
- Review OpenAI's Preparedness Framework to understand the criteria for different capability thresholds and associated safeguards.