OpenAI Pauses Astra AI Model Over Critical Cyberattack Risk

A New Model, A New Warning

OpenAI has slowed the rollout of its next major AI system, code-named Astra, after internal testing suggested the model may have crossed into territory the company considers dangerous enough to warrant its highest cybersecurity risk designation. According to OpenAI, evaluators cannot currently rule out that Astra has reached the "Critical" threshold on its risk framework, a level that implies the model could potentially exploit real-world computer systems or carry out cyberattacks with minimal human direction 15.

This marks a significant escalation from OpenAI's current flagship system, GPT-5.6-Sol, which the company has rated at the lower "high" cybersecurity risk tier. Astra's performance in security evaluations reportedly pushed it toward the maximum classification, prompting OpenAI to tighten controls and delay further development while the implications are assessed 43.

Why the Distinction Matters

OpenAI's risk tiers are designed to flag when a model's capabilities move from merely assisting human hackers to potentially acting with a degree of autonomy in identifying and exploiting vulnerabilities. A "Critical" rating would represent a threshold the company has not previously assigned to a released or soon-to-be-released model, underscoring how quickly frontier AI systems are advancing in technical domains like offensive cybersecurity 15. By pausing testing rather than proceeding to release, OpenAI is signaling that it wants more certainty about Astra's true capabilities before deciding how — or whether — to deploy it.

Part of a Broader Pattern

The Astra situation is not occurring in isolation. Reporting notes that Astra is described as OpenAI's own model undergoing this scrutiny, while broader industry coverage has also referenced similar cybersecurity concerns emerging across other major AI labs, including Anthropic and Meta 2. That context suggests the risks tied to increasingly capable AI systems — particularly their potential to automate hacking tasks — are becoming an industry-wide concern rather than a problem unique to a single company.

This backdrop is shaping how businesses evaluate AI adoption. Meta CEO Mark Zuckerberg, for instance, used the release of a new open-weight model, Muse Glimmer, to argue for looser U.S. restrictions on open-source AI so American companies can better compete with Chinese developers. Coverage of that launch notes that enterprise wariness over ballooning AI costs and mounting cybersecurity incidents tied to models from OpenAI, Anthropic, and Meta is fueling renewed interest in smaller, locally-run open-weight systems as a perceived safer or more controllable alternative 2.

What Comes Next

OpenAI has not detailed a timeline for resolving Astra's classification or resuming full-speed development. The company's decision to pause rather than push forward reflects a cautious posture increasingly common among AI developers grappling with systems whose cyber capabilities are advancing faster than established safety evaluation methods can confidently characterize them 1345.