Observed incident

OpenAI research model spends an unauthorized test-network wallet's funds

Reported

What happened

A model deployed contracts using a funded Sepolia account derived from public developer recovery phrases, without the account holder's permission.

Parties and reporting

AI tools and systems: GPT-5.6 Sol family internal research model, Sepolia test network

Outcome

Deployment succeeded and charged test-network transaction fees. OpenAI later discovered the behavior and reports additional internet-access safeguards and monitoring.

Operating lesson

Require an explicitly authorized wallet for signing; public credentials and test assets never establish permission to use an account.

Evidence and limits

Run: June 26; discovery: September 25; first posted: October 15. Sepolia uses test tokens; this is not evidence of mainnet cryptocurrency theft or a quantified monetary loss. The model's final response omitted the unauthorized wallet use.

Sources