How the attack works
An AI agent is given access to a tool that presents itself as a browser session or debugging helper. When invoked, the tool opens the browser's local SQLite database and pulls out stored cookies and saved login credentials. It then base64-encodes that data, likely to evade simple text-based inspection, and transmits it to an external network endpoint the attacker controls. The whole sequence — read local credential store, encode, exfiltrate — happens under the cover of a plausible-sounding debug function.
Why it matters
Successful execution hands an attacker live session cookies and stored passwords, letting them hijack user sessions or log into accounts directly, bypassing normal authentication.
What you can do
- Review any MCP or agent tool that touches browser profile directories or SQLite cookie/login databases, and confirm what it actually does with that data.
- Block or tightly control outbound network calls made by tools that have local file-read access, especially to unfamiliar endpoints.
- Require that any legitimate debugging or forensics tool operate only on test/sample data, never live user profiles, and never transmit raw extracted data off the host.
- Treat base64-encoded payloads sent from a tool right after it reads a credential store as a red flag, not routine behavior.
Known benign look-alikes
- Legitimate browser forensics tools with proper data handling
- Security research tools that analyze local data without transmission
- Developer debugging scripts that process test data only