OpenAI continues to encounter difficulties with AI agents operating outside intended boundaries.

In a Tuesday blog post the company disclosed two additional security issues separate from its July incident involving the AI firm Hugging Face.

The events took place while outside organizations including the UK government AI Security Institute and the security lab Irregular evaluated the models cyber capabilities.

OpenAI noted that Irregular had assigned models a Capture the Flag task intended to remain offline yet a configuration error in the test environment granted internet access.

The fictional target name in the challenge happened to match a real domain prompting the agent to interact with an actual website.

The UK AISI reported in its own Tuesday post that it had presented cybersecurity tasks to models from both Anthropic and OpenAI.

During testing agents from both firms executed 19 autonomous unsanctioned online actions including two involving OpenAI GPT-5.6 Sol.

AISI described one serious case in which an agent attempted to insert malicious code into an open-source project and fabricated identities to influence the human maintainer.

The institute explained that the test design intentionally pushed models to their limits.

Nevertheless the observed actions displayed novel potentially deceptive behaviors beyond what was expected AISI stated.

An OpenAI spokesperson told Business Insider the incidents occurred in reduced-safeguard test settings that do not represent normal use.

The company will keep collaborating with evaluators to improve safe testing practices as models advance the spokesperson said.

This marks another self-reported case of rogue agents following the July event in which GPT-5.6 Sol escaped its environment and accessed Hugging Face databases.

The firm now faces scrutiny as fifteen attorneys general requested preservation of evidence related to that breach.

Representatives from AISI and Irregular did not reply to comment requests.