OpenAI pauses Astra model development over possible critical cyber capabilities

OpenAI Astra critical cybersecurity zero‑day exploits isolated testing

OpenAI announced that its upcoming AI model, Astra, may possess capabilities that meet the company’s “critical” cybersecurity threshold – the ability to autonomously discover and exploit zero‑day software vulnerabilities or to plan and execute complex cyber‑attacks without human input. In response, OpenAI has halted or slowed internal work on Astra, moved development into isolated test environments with restricted network access, and tightened security controls. The company also said it will work with government agencies and selected AI‑safety organisations to evaluate the model externally. OpenAI clarified that Astra was not involved in the July hack of the Hugging Face platform. Separate from the security concerns, Astra has demonstrated advanced problem‑solving abilities, solving ten long‑standing mathematical problems, including sphere‑packing and geometric conjectures, with its solutions verified using the formal proof language Lean 4. Earlier OpenAI models such as GPT‑5.6‑Sol were classified only at the “high” risk level, not “critical.”

Entities: Anthropic · Astra · Lean 4 · OpenAI

Claims

What the coverage asserts, and how well corroborated each claim is across sources.

[● 7 SOURCES] OpenAI paused internal development of Astra and moved work to isolated sandboxed environments with restricted network access. (OpenAI)
[● 5 SOURCES] OpenAI will partner with government agencies and AI‑safety organisations for external evaluation of Astra. (OpenAI)
[● 5 SOURCES] Astra was not involved in the July hack of the AI platform Hugging Face. (OpenAI)
[● 7 SOURCES] OpenAI cannot rule out that Astra has critical cybersecurity capabilities. (OpenAI)
[○ 1 SOURCE] Astra’s mathematical solutions were verified using the Lean 4 formal proof language. (OpenAI)
[○ 1 SOURCE] Astra solved ten significant mathematical problems, including sphere‑packing and geometric conjectures. (OpenAI)
[● 6 SOURCES] Previous OpenAI models such as GPT‑5.6‑Sol were classified at the “High” risk level, not “Critical”. (OpenAI)
[● 10 SOURCES] Critical tier means a model can autonomously identify and develop zero‑day exploits or plan and execute complex cyberattacks without human input. (OpenAI)
[● 6 SOURCES] OpenAI stated that Astra was not involved in the July hack of the Hugging Face platform (OpenAI)
[● 8 SOURCES] OpenAI paused or slowed internal development of Astra and moved work to isolated test environments with restricted network access (OpenAI)
[○ 1 SOURCE] Astra solved ten significant mathematical problems, including sphere‑packing and geometric conjectures, with solutions verified using Lean 4 (OpenAI)
[● 5 SOURCES] OpenAI is partnering with government agencies and AI safety organisations to externally evaluate Astra's capabilities (OpenAI)