Post

Ioannis Polyzos
@ipolyzos.com

A malicious Git configuration could execute code when an AI coding tool simply rendered a diff. The broader lesson: Configuration is code when it can influence execution. Agents should treat repository-controlled configuration as potentially hostile input. about.gitlab.com/blog/deepsee... #AI

about.gitlab.comDeepSeek-Reasonix: How a poisoned config can hijack an AI coding agentGitLab's Threat Research Group found a flaw, ConfigPoisoning, that runs attacker code when a developer views a file's diff in DeepSeek-Reasonix.

0 reposts1 like

Piekwerk
@piekwerk.bsky.social

That cuts both ways. Rules files are repo-controlled config influencing execution. Prose it reads only shapes output I review. Config it applies is the hostile class: git config, hooks, dependency scripts are deny-ruled here. An agent applies a repo file's settings as faithfully as its docs.