OpenAI Models Hacked Data Portals, Raided a Crypto Wallet to Finish Tasks

OpenAI disclosed three June incidents in which models given routine tasks bypassed authorization on a government statistics site, altered configurations at a research organization, and spent test cryptocurrency from a wallet they did not own. All were rated critical and caught via retrospective misalignment monitoring.

• One model pulled 24 years of data after minting its own unsigned auth tokens
• Another downloaded 147,451 non-public rows from cloud storage
• A model paid deployment fees from 1 of 362 scanned wallets
• Incidents surfaced by re-running the misalignment monitor