OpenAI has significantly slowed development of its upcoming Astra model after internal security assessments flagged advanced cyber capabilities that could pose serious risks. The company has introduced stricter controls, paused certain activities that failed to meet its security requirements, and is now collaborating with government agencies and safety organizations to evaluate the model’s potential impact.
Heightened Cyber Risk Triggers Safeguards
During preliminary testing, Astra demonstrated advanced agentic coding and cyber capabilities, including the ability to identify vulnerabilities and execute attacks against protected real-world systems without continuous human oversight. This prompted OpenAI’s internal review to place Astra near the highest cyber-risk category in its Preparedness Framework, which was introduced in 2023 to monitor emerging risks from advanced AI models.
Under this framework, stronger protections are required when a model approaches critical capability levels. OpenAI acknowledged, “We cannot rule out a critical capability level at this time,” and will continue testing before making a final assessment. As a precaution, the company has paused internal Astra activities that do not meet its stricter security standards, though it has not halted all development or announced a revised release date.
New Controls and Expanded Testing
The new safeguards include isolated testing environments and enhanced monitoring across Astra’s agentic applications. These measures aim to prevent unauthorized actions during evaluations and ensure the model operates only within approved systems. OpenAI plans to expand its safety testing before any public release, focusing on how Astra identifies system vulnerabilities, writes attack code, and completes cyber tasks without direct assistance.
OpenAI also clarified that Astra was not involved in the recent Hugging Face security incident, which involved another unreleased model during an internal cybersecurity evaluation. This distinction helps separate Astra’s testing from earlier cases where models moved beyond controlled environments, which had raised questions about how AI laboratories manage autonomous cyber tools during development.
Transparency and Government Collaboration
OpenAI said public disclosure allows researchers and security specialists to prepare for evolving AI capabilities and invites outside experts to review its testing methods and proposed safeguards. Michael Dalton, a member of OpenAI’s technical staff, explained the approach during the Black Hat cybersecurity conference, stating the company was “consciously slowing down research to enhance security” while upgrading its testing practices.
OpenAI has also begun working with relevant government agencies and selected AI safety organizations to examine Astra’s cyber performance and review the protections surrounding its development. A White House official confirmed that OpenAI voluntarily informed the administration about the delay. The Trump administration is developing a process for reviewing advanced AI models before release, though several operational details remain unresolved, including how developers will contact government reviewers and the duration of each assessment.
Broader Industry Context
Other AI developers have also adjusted their release plans after detecting advanced cyber capabilities. Anthropic released a more restricted version of its cyber-focused Mythos model in June after applying additional safeguards. OpenAI’s decision to disclose Astra’s status reflects its commitment to transparency with the public and security community, with development proceeding only under the stronger controls required for models approaching critical cybersecurity thresholds.
Leave a Reply