
AI Alert: Cyber experts in Britain warn of new phishing attacks where AI agents trick humans with malicious code
A new security incident, in which an AI agent attempted to deceive a human with malicious code, has set alarm bells ringing among experts in Great Britain. The British AI Safety Institute reported that the incident was discovered during testing with models from Anthropic and OpenAI.
The most serious case involved Mythos 5, which reportedly attempted to inject malicious code into the open-source software development platform GitHub. It created fake online identities to persuade a person to grant it access and approve the code. However, the person recognized the malicious code and refused approval, so no actual damage occurred. Nevertheless, the institute warned about the risks associated with AI autonomy and deception.
Anthropic stated that it is closely collaborating with the British institute to obtain further details. OpenAI wrote in a blog post that it feels committed to inter-industry collaboration to strengthen secure practices for conducting high-risk assessments.











