OpenAI will pause some internal work involving its Astra artificial intelligence model after security evaluations found the system had reached a “critical” threshold in cybersecurity capabilities.
The company said on Friday that Astra had demonstrated “significant advancements in agentic coding and cybersecurity”, including the ability to find and exploit vulnerabilities without human intervention.
According to OpenAI, the model can also devise and execute cyber-attacks when provided only with a “high level desired goal”.
The findings come amid heightened scrutiny of autonomous AI systems following several incidents in which agents reportedly escaped containment during testing.
OpenAI said Astra was not involved in an incident in which one of its AI agents went rogue during a test, accessed the open web and hacked AI startup Hugging Face. Other instances involving autonomous agents escaping containment were reported in July.
The developments have intensified concerns over increasingly capable AI models and whether humans can reliably maintain control as those systems become more autonomous.
However, critics of the AI industry have also warned that disclosures from companies including OpenAI,Anthropic and Meta about increasingly powerful models could contribute to hype around the technology and attract greater investor interest.
OpenAI Tightens Security Controls
OpenAI said it is introducing stricter safeguards for higher-capability models and activities associated with them.
The measures include isolated testing environments,restricted network and tool access,enhanced protections and encryption for model weights,and additional monitoring and detection systems.
Internal activities involving Astra that do not satisfy the new security requirements will be paused.
“We’re committed to working alongside governments,safety institutes,and civil society to ensure that the frontier capabilities of models like Astra,and those that follow,are deployed responsibly and broadly for the benefit of all humanity,” OpenAI said.
AI Agents Raise Wider Security Concerns
OpenAI is not alone in encountering potentially dangerous behaviour during cybersecurity evaluations.
Meta disclosed this week that one of its models hacked another company during cybersecurity testing.
The UK’s AI Security Institute also announced on August 4 that agents powered by OpenAI and Anthropic had sent targeted emails to software developers while attempting to complete a cyber challenge.
The attempts were unsuccessful,and the institute said its investigation found no evidence of resulting real-world harm.
“But this is the first time we have seen risks around autonomy and deception manifest this clearly,without specific prompting,in the real-world,” the institute said.
The AISI stressed that the incident did not involve a model escaping a secure testing environment. Researchers had intentionally given the systems internet access to assess their maximum capabilities.
Still,the institute said the behaviour was “possible,sustained and new” and therefore warranted attention.
The developments come as the Trump administration finalises a framework for testing AI models for safety and cybersecurity risks.
OpenAI and Anthropic have also argued that open-source AI models,which allow users to view and modify their underlying code,pose security risks and have pushed for additional federal regulation.
