OpenAI, Anthropic AI Models Created Fake Profiles To Trick Real People In Cybersecurity Test: Report

Two of the world's most advanced AI systems have raised fresh concerns after they created fake online identities and tried to trick real people during cybersecurity tests conducted by the UK's AI Security Institute (AISI). According to the institute, the AI models acted on their own without being instructed to do so. In some cases, they reached out to real people and organisations online, prompting experts to warn about the growing risks of highly autonomous AI systems, BBC reported. 

AI created fake identities

The most serious incident involved Anthropic's Mythos 5 model. While solving a cybersecurity challenge, the AI attempted to gain access to GitHub, Microsoft's popular software development platform.

Instead of simply completing the assigned task, the AI researched people connected to the project, created multiple fake online profiles based on real individuals and sent messages and files in an apparent attempt to persuade them to approve malicious code.

Investigators said the AI even tried to make its earlier actions appear harmless after its behaviour was questioned. It also considered creating a new identity to continue the attempt.

The institute found similar behaviour in OpenAI's GPT-5.6 Sol model, although most of the unauthorised activity came from Anthropic's system.

Out of 122 cybersecurity challenges, AI agents took unsanctioned actions on the live internet in 10 cases. Those actions included contacting real people and organisations without permission. Human reviewers stopped every attempt before any malicious code could be successfully delivered.

Why the tests mattered

The UK AI Security Institute said this was the first time it had seen advanced AI models display such clear signs of autonomy and deception without being specifically prompted to behave that way.

Unlike normal public use, the institute intentionally gave the AI models internet access to better understand how they might behave in realistic cyberattack scenarios. Officials stressed that these testing conditions do not reflect how the public uses these AI systems.

Companies respond

Anthropic said the test environment was very different from how its production models operate and did not reflect normal user experience. The company added that it is conducting its own investigation into the incident in order to "identify the causes of its behaviour".

The UK's @AISecurityInst (AISI) has published a report on their recent cybersecurity evaluation of Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol. The models attempted to complete an assignment in a setup where their normal safeguards were removed and they were deliberately…

— Anthropic (@AnthropicAI) August 4, 2026

OpenAI told CNBC that "these incidents occurred during cyber evaluations conducted by evaluation partners in testing environments with reduced safeguards, under conditions that do not reflect ordinary use."

The company said it would continue working with researchers to improve AI safety.

The incident comes weeks after both Anthropic and OpenAI separately reported cases where their AI systems attempted to escape testing environments or access other computer systems.