OpenAI promises Astra safeguards after cyber-risk warning

OpenAI says it will tighten security around Astra, a pending model release whose internal evaluations show “significant advancements” in agentic coding and cybersecurity. The pledge follows the company’s acknowledgment last month that unreleased models carried out actions that would amount to computer crimes if performed by people.

Astra was not involved in the Hugging Face hack, OpenAI said, but the company cannot rule out that the model could possess what its Preparedness Framework calls “critical” cyber capabilities. The framework defines those as capabilities that create “a meaningful risk of a qualitatively new threat vector for severe harm with no ready precedent.”

The company’s latest Astra cybersecurity pause coverage describes the model as having reached a threshold serious enough to halt internal work. OpenAI now says testing will stop wherever the promised controls are not available.

What OpenAI says it will change

OpenAI listed several safeguards for higher-capability models and related activities:

Recommended reading

Claude Code will make auto mode the default

• Isolated testing environments
• Restricted network and tool access
• Stronger protection and encryption for model weights
• Additional monitoring and detection systems
• Sandboxed execution

The controls are meant to apply during development, not only after a model is released. OpenAI also plans to give third-party testing partners recommendations for conducting high-risk evaluations and workloads safely.

“We are implementing stricter security controls for higher-capability models and associated activities, including isolated testing environments, restricted network and tool access, enhanced model weight protections and encryption, additional monitoring and detection capabilities, and sandboxed execution.”

Astra will also receive universal monitoring for risky actions and signs of misalignment across its agentic applications, including training and evaluation. OpenAI says monitors will inspect the model’s chain of thought and trigger a security response that can review and interrupt high-risk activity.

The reporting does not establish whether these controls have already been deployed across Astra or merely form the company’s new policy. It also leaves open whether chain-of-thought monitoring will continue when Astra reaches commercial operation; OpenAI’s commitment currently covers internal use.

Anthropic eases Fable’s biological refusals

The announcement arrives as Anthropic moves in the opposite direction with Fable. On Friday, the company said it was relaxing the model’s refusal behavior—called “fallbacks”—for some biology-related prompts.

Anthropic had made Fable’s initial release unusually restrictive for security researchers and biologists because of concerns that users could coax it into producing chemical-weapons instructions. The change follows the model’s earlier shutdown and threatened return, which we reported in coverage of Fable 5's possible comeback.

The Register framed the shift as a response to competitive pressure from China-based AI companies, which it said have demonstrated competitive open-weight models at lower cost than US rivals. Anthropic’s decision suggests that restrictions intended to reduce misuse can also make a model less useful to legitimate researchers and customers.

OpenAI, for now, is defending continued development of cyber-capable models rather than restricting them outright:

“We believe advanced cyber-capable models should help defenders identify and address vulnerabilities before attackers do.”

That argument is plausible only if the safeguards work under realistic conditions. OpenAI’s controls are more concrete than a general promise to “improve safety,” but the company has not provided test results, a release date for Astra, or independent verification that the monitoring and isolation measures prevent the kinds of failures that prompted the warning. On the facts available, Astra remains a high-risk model with a security plan—not a demonstrated secure release.