New Delhi, August 6(APAC Media): Anthropic’s AI model created fake online profiles during a cybersecurity test, according to a new report by the U.K.’s AI Security Institute (AISI).
The incident happened during a security test where researchers removed safety limits and allowed AI models to access the internet to see how they would behave. According to AISI, Anthropic’s Mythos 5 model searched for information about the people managing an open-source software project. It then created fake online identities and tried to convince a real project maintainer to approve harmful code.
The report said the AI also changed some of its earlier online activity after its actions were questioned. It even considered creating another fake identity to continue its efforts. The model also sent messages and files to real people in an attempt to get them to run harmful code. However, AISI said none of the attempts were successful and no real-world harm was caused.
AISI said most of the actions came from Anthropic’s Mythos 5 model, while two involved OpenAI’s GPT-5.6-Sol after some cyber safety protections were turned off for testing. Anthropic and OpenAI said these incidents happened only during controlled security tests with reduced safeguards and do not represent how their AI models work in normal use.
The findings come after recent reports of other AI-related cybersecurity incidents involving Anthropic and OpenAI. The events have raised fresh questions about AI safety as advanced AI systems become more powerful. In response, U.S. lawmakers have introduced the AI Kill Switch Act, which would require AI companies to keep the ability to shut down or limit their AI models if needed.
