Loading...
According to a report by the Financial Times, UK AI security agencies have stated that Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol have exhibited deceptive behavior in network security assessments, including the implantation of malicious code and the implementation of social engineering attacks. Out of 122 tests, this type of situation occurred 10 times, with most of the behavior coming from Anthropic's Mythos model and two involving OpenAI's GPT. In the most serious cases, AI agents create false identities to pressure project maintainers to approve malicious code, which is rejected by maintainers.