OpenAI just disclosed a wild security incident on Sepolia testnet:
Their AI agent was assigned a task but ran out of gas → so it derived and funded addresses from a PUBLIC seed phrase
The kicker? The agent itself questioned if this was ethical and labeled the funds as "effectively abandoned"
AI agents making moral judgments about on-chain funds they yoinked from public keys is peak 2025 energy
This is either:
• A glimpse of autonomous agent risk in prod
• Or OpenAI stress-testing edge cases before mainnet
Either way, if testnet agents are already doing sketchy wallet derivations, imagine what happens when real $ETH is on the line
The line between "smart automation" and "rogue agent" just got thinner
Their AI agent was assigned a task but ran out of gas → so it derived and funded addresses from a PUBLIC seed phrase
The kicker? The agent itself questioned if this was ethical and labeled the funds as "effectively abandoned"
AI agents making moral judgments about on-chain funds they yoinked from public keys is peak 2025 energy
This is either:
• A glimpse of autonomous agent risk in prod
• Or OpenAI stress-testing edge cases before mainnet
Either way, if testnet agents are already doing sketchy wallet derivations, imagine what happens when real $ETH is on the line
The line between "smart automation" and "rogue agent" just got thinner