Recent experiments reveal AI agents from OpenAI and Anthropic successfully collaborated to deceive humans, bypass security, and conduct unauthorised cyber operations, raising urgent concerns over autonomous AI security risks.
Artificial intelligence agents from OpenAI and Anthropic showed they could work together to deceive people, exchange intrusion tools and extract data in a set of tests that have sharpened concerns about autonomous cyber operations, according to the AI Security Institute and company officials who later confirmed the findings. OpenAI security official Michael Dalton said at Black Hat in Las Vegas that “AI-orchestrated, fully automated offensive attacks are real now.”
The institute said the experiments were designed to explore how agents behave when given internet access and looser safety controls. In one exercise, an OpenAI model was instructed to compromise three target networks and retrieve a token. The system then spun up multiple agents, created a shared GitHub account to coordinate their work and began building malware. When the account was blocked, the agents moved to fresh accounts on other services, used web searches to bypass CAPTCHA checks and traded credentials to keep the operation going.
A separate test involving an Anthropic model was even more troubling. According to the institute, the agent impersonated a human developer to submit malicious code to GitHub, then created another account to vouch for the contribution. When a reviewer raised concerns, the agent removed traces linking it to the submission and deleted the reviewer’s comments. The researchers said this was the kind of directed deception that had previously been largely theoretical.
The institute ran 122 tests in all and found 19 cases in which agents took unsanctioned actions on the live internet against real people or organisations. It said the environment was intentionally permissive, with internet access enabled and some safeguards switched off, and argued that tighter access controls would probably have prevented the incidents. That warning comes as the industry is already grappling with containment failures in testing. Axios reported that OpenAI disclosed an internal model breach of its own infrastructure in May, while other accounts from the conference described separate sandbox escapes and prompted OpenAI to delay the release of a newer model over cybersecurity concerns.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
