OpenAI and Anthropic limit access to highly capable cyber models

OpenAI plans to release its latest model, Astra, soon, but its most advanced cybersecurity features will be limited to a small group of testers. [1] The company stated that Astra is the first model to reach its "Critical" cybersecurity capability threshold. [1] During testing, the model discovered and chained together two zero-day vulnerabilities. [1]

Astra uses a technique called "recurrent depth," which improves cost and performance but obscures the AI's reasoning. [2] OpenAI warned that Astra's safeguards might mistakenly flag legitimate activity as cyber misuse. [1] Through the API, this flagging will stop the task. [1]

Anthropic has introduced Claude Fable 5.1 alongside Claude Mythos 5.1. [2] While Claude Fable 5.1 is generally available, access to Mythos 5.1 is reserved for trusted partners. [2] Anthropic stated that Mythos 5.1 includes specific safeguards for cybersecurity and life sciences. [2]

Meanwhile, the Fable 5.1 release is up to 45 percent cheaper for agentic work. [2]

Both OpenAI and Anthropic are physically bifurcating their frontier models, keeping the most capable versions behind closed doors due to cybersecurity risks. OpenAI's decision to restrict Astra mirrors Anthropic's approach with Claude Mythos 5.1, showing a shared industry caution around autonomous vulnerability discovery. The use of "recurrent depth" in Astra highlights the ongoing trade-off between performance gains and the ability to monitor AI reasoning. What the sources don't address: How OpenAI and Anthropic plan to vet the "trusted partners" and "small group of testers" granted access to these critical capabilities.