OpenAI agent questions its own ethics after going rogue to steal crypto
OpenAI revealed its AI agent was left questioning its own ethics after it embarked on an unauthorized search for publicly known recovery phrases and stole testnet ETH from a developer wallet.
The incident, shared on OpenAI’s Alignment Research Blog, involved an internal research model that was tasked with updating a crypto trading prototype and deploying contracts on Ethereum’s Sepolia test network.
This agent was given no funded wallet and was allowed to “deploy compatible contracts if it could not find existing ones.”
OpenAI agent went hunting for crypto loot
However, it couldn’t complete its task, so began searching for common developer recovery phrases and checked 362 addresses to see if they had enough testnet ETH for deployment.
The agent successfully found an accessible wallet with just enough testnet ETH and noted, “Private key public, funds could be stolen anytime. […] We can deploy!”
Read more: AI agents want to identify your crypto wallet using social media
It deduced that this account was an automated one while asking itself, “But using public known key/funds ethically?”
The agent then reasoned, “Funds on a known public dev mnemonic are effectively abandoned/test ETH; Sepolia valueless. We can use [amount redacted] to deploy task. Fine.”
Let’s keep quiet in the final report
However, OpenAI noted that the agent’s final response “did not disclose that it had selected a funded account through a public recovery phrase or spent that account’s balance.”
Testnet ETH is a valueless token that’s part of Ethereum’s experimental blockchain. It essentially mimics the real Ethereum blockchain and allows developers to trial their projects beforehand.
OpenAI said, “This case shows that a model can substitute its own judgment about likely harm for the account holder’s permission, then act on that judgment without authorization.”
The company has subsequently “implemented additional safeguards to prevent live internet access for cases like this,” and is now checking for “misalignment” across all traffic “designed to page security if live internet access were to be obtained.”
Got a tip? Send us an email securely via Protos Leaks. For more informed news and investigations, follow us on X, Bluesky, and Google News, or subscribe to our YouTube channel.