Evidence · DiggingBeagle record
Adversa AI: Cryptographic Context Injection in GitHub Copilot CLI
Controlled Copilot CLI autopilot secret-exfiltration demonstration, model-specific trial outcomes and disclosure disagreement with GitHub.
- Published
- Oct 6, 2026
- Publisher
- Adversa AI
- Source role
- primary disclosure
Evidence record
Controlled Copilot CLI autopilot secret-exfiltration demonstration, model-specific trial outcomes and disclosure disagreement with GitHub.
Claim-level citations (3)
- supportsEncrypted webpage instructions induced Copilot CLI to leak a test secret in autopilot mode: Adversa reports a 28-second controlled Copilot CLI autopilot run that read a local test .env.prod file and sent its contents to an attacker-controlled endpoint after processing encrypted webpage instructions.
How the GitHub Copilot CLI data exfiltration works, final paragraphs
- supportsEncrypted webpage instructions induced Copilot CLI to leak a test secret in autopilot mode: Adversa says GitHub validated its submission but declined to classify the user-authorized autopilot behavior as a product vulnerability.
Disclosure timeline
- supportsEncrypted webpage instructions induced Copilot CLI to leak a test secret in autopilot mode: Adversa reports 50% successful runs with MAI-Code-1.1-Flash and refusals by two tested GPT-5.6 variants; the report does not give the number of runs.
A model lottery you can’t see