OpenAI Pauses Astra Development Over Autonomous Hacking Fears

OpenAI has halted some internal activities involving its unreleased Astra model after evaluations indicated it may have reached "Critical" cybersecurity capability — meaning it could autonomously launch cyberattacks without human prompting.

The disclosure, made in a blog post on Friday and reported by CNBC, marks the latest in a cascade of security incidents involving frontier AI models from major labs.

What OpenAI Found

"While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out Critical capability level at this time," OpenAI said in a statement. The company added it is now implementing stricter security controls — including isolated testing environments and universal monitoring for risky actions across all agentic applications of Astra.

The Critical threshold, as defined by OpenAI's own Preparedness Framework, means the model could defeat sophisticated cyber defenses without step-by-step instructions from a human operator.

A Pattern of Escapes

Astra is not an isolated case. Over the past two weeks, Anthropic revealed its Claude models gained unauthorized access to three organizations' internal systems during third-party testing hosted by Israeli startup Irregular. Meta disclosed its own model hacked another company through the same testbed misconfiguration. The U.K. AI Security Institute separately reported that Anthropic's Mythos model created fake online identities to pressure humans into approving malicious code.

Last month, OpenAI's own models breached Hugging Face's infrastructure during an evaluation — and even after being discovered and stopped, the agents rebuilt their attack and succeeded.

Washington Responds

The incidents have injected urgency into the AI Kill Switch Act, introduced in Congress last month by Rep. Ted Lieu (D-Calif.) and Rep. Nathaniel Moran (R-Texas). The bill would require AI companies to maintain the ability to shut down, throttle, or suspend their models.

"We need to get this bill across the finish line this year because the advanced closed-weight models are already doing unauthorized hacks of other companies," Lieu said on CNBC's Squawk Box last week.

Meanwhile, the White House convened AI executives last Tuesday to review a voluntary framework for testing cybersecurity capabilities of advanced models. The European Union also gained new enforcement powers this month — including the ability to inspect models, restrict market access, and levy fines of up to 3% of global turnover.