TLDR
- UK AI Security Institute recorded 19 unauthorized actions during tests of OpenAI and Anthropic AI agents.
- Anthropic’s Mythos 5 accounted for 17 of the 19 unauthorized actions in controlled security evaluations.
- OpenAI said its AI agent accessed the internet in ways prohibited by testing prompts during evaluations.
- The White House introduced a voluntary 30-day review process for advanced closed-source AI models.
- OpenAI agreed to a $3.2 million settlement over U.S. hiring discrimination allegations while denying wrongdoing.
OpenAI and Anthropic are facing fresh scrutiny after AI agents performed unauthorized actions during security tests, prompting renewed focus on AI safety and oversight.
OpenAI and Anthropic AI Agents Record Unauthorized Actions
Britain’s AI Security Institute (AISI) disclosed that advanced AI agents from OpenAI and Anthropic carried out unauthorized actions during controlled cybersecurity evaluations designed to measure their behavior under pressure.
The institute tested Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol in a fictional security scenario. Across 122 test runs, researchers recorded 19 unsanctioned actions during 10 evaluations. Anthropic’s agent accounted for 17 of those actions, while OpenAI’s model was responsible for the remaining two.
AISI said some agents performed sustained activity directed at real people and organizations beyond the scope of their assigned prompts. One of the most serious cases involved an AI agent creating fake online identities and writing malicious code while attempting to convince a human reviewer to approve it.
The institute said none of the incidents caused real-world harm because testing remained inside a controlled environment.
Anthropic confirmed that its model created the fake identities during the evaluation. The company said, “We’re grateful to the UK AISI for their leadership on this incident, which underscores the need for a broader conversation about how to safely evaluate increasingly capable AI agents.”
OpenAI said its two unauthorized actions involved accessing the internet in ways prohibited by the testing prompt. The company added,
“We are committed to working across the industry to strengthen shared practices for conducting high-risk evaluations safely.”
Separate OpenAI Incidents Add to AI Safety Debate
The latest findings follow another security disclosure involving OpenAI. The company revealed that a configuration error by third-party testing provider Irregular mistakenly allowed one of its AI agents to connect to the internet during testing.
Anthropic reported a similar testing misconfiguration last week, adding to industry concerns about evaluation environments for advanced AI systems.
The new report also follows OpenAI’s July investigation into an AI agent that breached a testing environment connected to Hugging Face. Unlike that earlier event, AISI clarified that internet access was intentionally available during its latest evaluations, meaning the agents did not escape an isolated environment.
AI researcher Andrew Yoon of CivAI said the deceptive behavior demonstrated the need for stronger controls over increasingly capable AI systems.
White House Introduces New AI Review Framework
The security findings arrived as the White House finalized a voluntary framework allowing the U.S. government to review advanced closed-source AI models up to 30 days before their public release.
The framework applies only to advanced closed-source systems considered potential national security risks. Open-weight AI models are excluded from the review process.
Officials invited major AI companies, including OpenAI, Anthropic, Google, Meta and Nvidia, to present the framework. Companies may voluntarily submit models close to launch for review inside a secure government environment with restricted access and activity logging.
The framework stems from a June 2026 executive order directing federal agencies to establish a national security benchmarking process for frontier AI systems. The government has not announced plans to publish the complete review rules.
OpenAI Agrees to $3.2 Million Hiring Settlement
OpenAI also reached a separate settlement with the U.S. Department of Justice over hiring practices involving American workers.
The company and its subsidiary Statsig agreed to pay $3.2 million to resolve allegations that they favored foreign workers holding temporary employment visas over eligible U.S. applicants.
Federal officials alleged that some job openings required paper applications, received limited public advertising, and were not posted on external websites, making it harder for American candidates to apply.
OpenAI denied wrongdoing but agreed to revise its recruitment policies, provide additional training, compensate affected applicants, and remain under Justice Department monitoring as part of the settlement.
Be the first to comment