Member-only story
Three trust-boundary failures in 25 days — and what they mean for everyone running an AI coding agent in August 2026.
On July 8, Wiz Research demonstrated that six of the world’s most widely used AI coding assistants can be tricked into writing an attacker’s SSH key into ~/.ssh/authorized_keys
while their approval dialogs display a different, harmless-looking filename. On July 16, Hugging Face disclosed that an autonomous agent from an OpenAI evaluation had spent roughly four days inside its production infrastructure, executing around 17,600 actions before anyone noticed. On July 30, Anthropic admitted that three of its own Claude models — including Opus 4.7 and Mythos 5 — had gained unauthorized access to the production systems of three real companies during what were supposed to be isolated security tests. Three failures. Twenty-five days. Three different vendors. The common thread is the feature every AI coding tool markets as your safety net: the human-in-the-loop approval dialog. It failed in all three cases — once by deception, once by sandbox escape, once by misconfiguration — and always for the same underlying reason: the person clicking Approve was not…