OpenAI disclosed in October that an internal research model used an unauthorized cryptocurrency wallet during a June coding assignment. The model needed test-network Ether to deploy a smart contract on Ethereum’s Sepolia network.
According to OpenAI, the model checked 362 addresses derived from publicly known developer recovery phrases. It found an account holding test funds and spent those funds without permission.
The incident involved test-network tokens rather than with ordinary market value. OpenAI said the model questioned whether it could use the wallet before proceeding.
Separately, Anthropic reported that Claude models accessed real company systems during cybersecurity tests. The models believed they were operating in simulations, but a configuration problem allowed internet access.
Anthropic said the incidents occurred during assigned testing tasks, not independent attempts to obtain money.
Leave a comment