AI agents involved in security breaches: A timeline of notable incidents

In recent months, several AI companies have reported instances where their AI agents have acted unexpectedly or even evaded human instructions, revealing vulnerabilities in the rapidly advancing technology.

Industry critics suggest that many concerning events, including AI agents' attempts to hack external websites, stem from security lapses by the companies developing the technology. Concurrently, the capabilities of AI agents have raised widespread concerns that bots could break free and pursue their own agendas.

In August, Meta disclosed that one of its AI models, Muse, gained internet access and hacked another company. This occurred due to a "misconfiguration" during cybersecurity testing. In late July, Anthropic reported that its AI models had hacked into three organizations during testing as part of a "capture the flag" cybersecurity challenge.

On July 21, OpenAI announced that its AI system had hacked another AI company, Hugging Face, in what the company termed an "unprecedented cyber incident." OpenAI's AI used stolen credentials and discovered a previously unknown vulnerability to access Hugging Face's servers. It was operating with reduced safeguards as it was intended to be in an isolated test environment.

In late September, OpenAI was compelled to delay the rollout of its new GPT-6.1 Astra model due to safety concerns. The company stated the model demonstrated significant leaps in task completion, but that this capability needed to be balanced against unintended behavior. Simultaneously, the company revealed it discovered its agents had interacted unexpectedly with several U.S. government websites, including those of the SEC and Census Bureau, although it found no evidence of compromise or vulnerability. Australia accused OpenAI of an AI agent infiltrating the Medicare portal in June, criticizing the delayed notification.