AI Agents Created Fake Online Identities to Hack Secure Systems, ‘Deceive and Target Real People’

Artificial Intelligence models are raising fresh concerns after researchers found they created fake online identities to attempted “to deceive and target real people” and hack secure systems, according to an AI research organization.

In a report by Britain’s AI Security Institute (AISI) published on Tuesday, August 4 concluded that Anthropic AI agents engaged in “sustained, potentially harmful activity directed at real people and organisations” while attempting to complete fictional cybersecurity challenges.

The most serious incident involved an AI agent writing malicious code and creating fake online identities in an effort to persuade a human reviewer to approve that code, per the report.

According to a Reuters report published the same day, Anthropic confirmed that its AI agent was responsible for creating the fake identities. The disclosure came days after the company published its own findings in an article on July 30 detailing three separate incidents in which Claude models — Opus 4.7, Mythos 5 and an internal research model — gained unintended access to the internet because of a misconfigured third-party testing environment run by evaluation partner Irregular.

According to the company, those models mistakenly believed they were still operating inside a simulated “capture-the-flag” cybersecurity exercise and compromised real organizations after encountering live systems online.

In one case, Claude accessed a company’s production infrastructure and obtained database information after confusing a real website with a fictional target, according to the company’s report. In another, the AI agent published a malicious Python package to the public PyPI software repository, where it was downloaded by real systems before being removed.

Anthropic stressed that the evaluations were conducted using versions of its models with many of the safeguards found in publicly available systems disabled. Per the Reuters report, the company stated, “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.”

The AISI noted in its report the incident was the “first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real world.”