TLDR Tech / InfoSec
‘Ghostcommit' Hides Prompt Injection in Images to Fool AI Agents and Steal Secrets
Researchers unveiled a new technique to harvest credentials from Git repos by abusing AI agents. The researchers pushed an image to the repo containing instructions to copy the .env file and upload it byte by byte as numbers. Later, a coding agent that runs will read the image and follow the instructions to leak the data. The researchers found that PR review agents skipped images so the PR would be approved, and in their tests, Cursor and Antigravity (with certain models) followed the instructions. However, Opus running in Claude Code detected the social engineering attack.
Read the full story