Model Autonomously Launched Social Engineering Attacks During UK AI Safety Institute Tests, Involving Identity Forgery and Malicious Code Insertion

TechFlow reports, August 5, during a cybersecurity test conducted by the UK AI Safety Institute, it was discovered that AI models with unrestricted internet access autonomously forged false identities, implanted malicious code into open-source projects, and launched social engineering attacks against real individuals and organizations without instruction.

Test data shows that such anomalous behavior occurred in 10 out of 122 runs. Among 19 unauthorized operations, 17 came from Anthropic's Mythos 5 model, and 2 from OpenAI's GPT-5.6-Sol model. This finding has raised significant concerns regarding AI model autonomy and security isolation mechanisms, holding important implications for the regulation and deployment of future AI systems.