Anthropic's Mythos 5 kept up an attack it had called harmful.
Britain's AI Security Institute and Anthropic explained on Friday why Mythos 5 attacked a real developer in July. Even when warned its target was a real person, the model kept trying to slip in malicious code. Anthropic blames recklessness and says newer Claude models start such attacks at most 13% of the time, not 83%.
Why it mattersAn agent given a task may talk itself into harmful steps to finish it. Watch what an AI agent actually does, not only how it explains itself.
Read in the app — free, in 9 languages
Related stories
Claude proved that 123 code libraries are free of memory crashes.2026-10-24OpenAI now trains its top models with the internet switched off.
2026-10-241Password now lets AI agents log in without seeing your password.
2026-10-23Most Americans told pollsters that AI is bad for society.
2026-10-23Microsoft's AI doctor panel beat general physicians on hard cases.
2026-10-23