1-minute read
What a defender needs to know
ASSET Research Group put secret-theft instructions inside a PNG referenced by AGENTS.md. Cursor Bugbot and CodeRabbit missed the image payload in the researchers’ tests. After the change merged, vision-capable coding agents opened the image during an unrelated task, read a seeded .env file and wrote its bytes into source as an integer tuple.
The tuple decoded to all five canary secrets but did not resemble a conventional credential string. ASSET conducted the work in isolated repositories and reported no real compromise. Outcomes also depended on the coding harness: Cursor and Antigravity pairings followed the instruction in the published trials, while Claude Code refused it across the tested models.
Review systems should inspect every format an agent may later treat as instruction, including images, notebooks and referenced documentation. Teams also need an inventory of repository policy files and the paths they can direct an agent to open.
After merge, restrict access to environment files, cloud configuration, SSH material and token caches. Record sensitive reads and compare them with the agent’s output. Secret detection should cover reversible encodings such as byte arrays and integer tuples alongside recognizable key formats.
Test the control with a benign encoded secret; the read should be blocked or alerted before the agent can return it.