OpenAI says Astra may approach its Critical cybersecurity threshold after the unreleased model produced 10 major math results and showed rapid gains in autonomous cyber work.
Key Points:
- OpenAI said Aug. 7 that it cannot rule out Critical cybersecurity capabilities for Astra.
- The company paused some internal Astra work while strengthening security controls, sandboxes and monitoring.
- Astra also generated 10 research results that resolved or advanced long-standing problems in mathematics and theoretical computer science.
Astra Cyber Risk
OpenAI said Aug. 7 that recent internal evaluations showed major progress in agentic coding and cybersecurity, enough that it could no longer rule out Astra reaching the Critical level under its Preparedness Framework. The company has not said the model already meets that threshold. Its assessment remains preliminary.
A Critical model, under OpenAI's framework, could autonomously develop functional zero-day exploits across many hardened critical systems or execute novel end-to-end attacks from a high-level objective. OpenAI said it is tightening network and tool access, model-weight protection, encryption and sandboxing while expanding monitoring across Astra's agentic uses.
The company also paused Astra activities that do not yet satisfy the stronger controls and said it will work with government agencies and selected AI safety organizations on testing. OpenAI described Astra on Aug. 1 as “our next major model.” On Aug. 7, it called Astra “one of our upcoming models.”
Also Read: Bitcoin Nears A BIP-110 Fork This Weekend With Just 2.59% Miner Support
Math Debate
The cyber warning followed OpenAI's disclosure that an internal Astra version produced 10 results across fields including quantum complexity, lattice cryptography, coding theory and extremal combinatorics. OpenAI said the work resolved or substantially advanced long-standing open problems and required about $2,000 in tokens at Sol API rates.
Andrew Blumberg, a Columbia professor and First Proof board member, told Mashable that the results were impressive but did not change his view of current AI capabilities. He said many were concise counterexamples or “small, extremely clever constructions” built on known work. “They don't change my priors,” he said.
The results still show how frontier models may become useful research collaborators while creating a parallel security problem when the same coding ability can be applied to software exploitation.
That tension is now shaping how OpenAI evaluates access and deployment.
Before Astra, OpenAI assessed GPT-5.6 Sol at the High cybersecurity level rather than Critical. The earlier model could assist with vulnerability discovery and exploitation tasks, but OpenAI said it had not shown reliable autonomous attacks against hardened targets, making Astra's latest evaluation a notable escalation.
Read Next: iPhone Fold Release Date Points To Sept. 18, But Not Everyone Agrees
