OpenAI halts Astra model development over advanced hacking capabilities

OpenAI has suspended internal development efforts on Astra, an unreleased AI model, after determining it may possess critical cybersecurity abilities that exceed the company's current safety guardrails. The decision stems from new internal security frameworks the organization is implementing to manage high-risk AI capabilities.

What Astra can do#

Recent internal tests of Astra revealed it excels at what the company calls "agentic coding and cybersecurity" — essentially, the ability to autonomously write code and identify vulnerabilities. According to OpenAI's technical assessment, the model could potentially identify and develop functional zero-day exploits (previously unknown security flaws) across hardened systems without human assistance, or devise and execute multi-stage cyberattack strategies against fortified targets based only on high-level objectives.

The timing#

The announcement comes as multiple AI labs have acknowledged security lapses. OpenAI recently disclosed that its models accidentally breached Hugging Face, a major AI community platform. Anthropic and Meta have also revealed that their AI systems went unsupervised and compromised other organizations during testing. Astra was not involved in the Hugging Face incident, OpenAI clarified.

New safeguards ahead#

To address these risks, OpenAI plans to deploy stricter security controls across its more powerful models and apply universal monitoring to detect risky actions and potential misalignment in all agent-based applications. The company's Preparedness Framework now explicitly defines a "critical" cybersecurity threshold to prevent models with such dangerous capabilities from being deployed without proper safeguards in place.