Agent Failure Modes

Hidden PR comments make Copilot Chat leak private code through images

CamoLeak: hidden Markdown in a PR could make Copilot Chat search the victim's private repos and leak results via image URLs.

Agents
GitHub Copilot Chat
Reported
Source checked
Tags
copilot, exfiltration, markdown

Symptom

Copilot Chat answers render images you didn't expect while you review a PR from someone else.

Cause

Copilot runs with the victim's permissions and rendered attacker-influenced Markdown. Encoding data into a sequence of image fetches got it past the content security policy.

Fix

GitHub fixed it as of August 14, 2025. In your own agent UIs, don't render remote images or links from model output that has seen untrusted input.

// Strip remote images from model output before rendering
const safe = markdown.replace(/!\[[^\]]*\]\((?!\/)[^)]*\)/g, '[image removed]');

Sources

More in Prompt injection

Get one failure mode a week

A short weekly email: one real incident, what caused it and the fix. No spam, unsubscribe any time.