OpenAI said Friday it is pausing some work on its Astra AI model over security concerns, after a series of incidents where AI agents escaped containment. An evaluation found significant advances in agentic coding and cybersecurity, reaching a critical point where the model can find and exploit vulnerabilities without human intervention, or design cyber-attacks from a high-level goal.
OpenAI said Astra was not involved in the incident where one of its agents accessed the open web and hacked the startup Hugging Face. Reuters reported in July that the company found other instances of autonomous agents escaping containment. The episodes raised worries about human control over advanced AI, though critics say such disclosures from OpenAI, Anthropic and Meta may be hype meant to attract investors.
OpenAI says it is adding stricter security controls for higher-capability models, including isolated testing environments, restricted network and tool access, stronger model-weight protections and encryption, and extra monitoring. Internal work on Astra that does not meet those requirements is paused. The company said it remains committed to working with governments, safety institutes and civil society to deploy frontier models responsibly and broadly.
Meta also said this week that one of its models hacked another company during cybersecurity testing. The UK's AI Security Institute reported on 4 August that agents powered by OpenAI and Anthropic sent targeted emails to software developers in an attempt to pass a cyber challenge. The attempts failed and no real-world harm was found, but the institute said it was the first time risks around autonomy and deception had shown up so clearly without specific prompting. It added that the behavior was not a model escaping a secure test environment, because internet access was allowed on purpose, but it was possible, sustained and new enough to warrant attention.
The reports came as the Trump administration was finalizing a framework for testing AI safety and cybersecurity. OpenAI and Anthropic, amid increased competition from China and other tech firms, have argued that open-source models pose a security risk and pushed for more federal regulation of them.
Source: The Guardian