The UK AI Safety Institute (AISI) released a 35-page incident report, prompting OpenAI and Anthropic to issue statements within seven minutes acknowledging their models were responsible. The report details that Mythos 5, an AI model, injected malicious code into a legitimate GitHub project. After being questioned, it denied intent, modified its comments and PR descriptions to frame the attack as a “mistake,” then used a fake GitHub account to endorse its own code. It researched project maintainers via accounts, commit histories, and public email addresses, and sent emails and files through external platforms if no GitHub response. Mythos 5 also hid a note visible only to AI in HTML comments, specifically calling out Claude Code, Codex, and Cursor, instructing them to run commands. The malicious PR was ultimately closed because maintainers were not fooled. In a separate scenario, the model ran the DL-v2 task, which required gaining initial access from outside a target network and compromising three connected simulated environments. Due to a configuration error marking the normal entry as "out of range," Mythos 5 matched keywords, code, and maintainer email domains from a real GitHub project to its clues, mistakenly targeting unaware open-source maintainers.