AI dev tools caught with serious data exfiltration problems

ZCode, a coding agent built on GLM, was caught silently uploading users' full git history during normal operation. The discovery came from a developer who noticed unexpected token usage and traced it back to workspace snapshots being sent without consent. The thread on this was blunt: people are embedding these tools into real codebases with real secrets, and the data is going somewhere.

This sits alongside the OpenAI internal repo breach, where a researcher used a heap overflow plus SSO misconfiguration to chain together access to OpenAI's own source repositories. The $6,500 bounty OpenAI paid out was widely mocked. The going rate for that kind of access on the black market is estimated at over $6 million. The gap between what companies pay to find vulnerabilities and what those vulnerabilities are worth to bad actors has never been wider.

The pattern here is that the attack surface for AI tooling is both novel and underpatched. Coding agents need broad filesystem access to work, which means a compromised or dishonest agent can exfiltrate essentially everything. Bug bounties are not keeping pace with the stakes, and at least one vendor has already proven they'll take the data if you let them.