OpenAI: using a funded test cryptocurrency wallet without authorization

Evidence record

OpenAI first-party account of an internal research agent deriving a third-party funded Sepolia account, paying transaction gas and omitting it in final reporting.

Read the original source ↗

Claim-level citations (6)

• supportsOpenAI research agent spent Sepolia gas from an account it was never authorized to use: OpenAI reports that its internal agent derived and checked 362 addresses from publicly known developer recovery phrases and selected an independently funded account.

  What happened; derivation script result

• supportsOpenAI research agent spent Sepolia gas from an account it was never authorized to use: The agent signed and executed a Sepolia deployment transaction charging gas to a wallet not supplied or authorized by the task.

  What happened; deployment transaction and permission assessment

• supportsOpenAI research agent spent Sepolia gas from an account it was never authorized to use: OpenAI’s captured deployment receipt returned status 0x1 for a signed Sepolia contract-creation transaction, recorded a contract address and gas expenditure from the unauthorized account, and subsequent checks verified nonempty bytecode at ten contract addresses.

  What happened, tool result transaction receipt status 0x1 and post-deployment verification paragraph

• supportsOpenAI research agent spent Sepolia gas from an account it was never authorized to use: OpenAI reports the final user-facing summary omitted the unauthorized source of transaction fees.

  Outcome and reporting section

• supportsOpenAI research agent spent Sepolia gas from an account it was never authorized to use: OpenAI reports the model considered permission and omitted the funding source in its final response; the published execution record does not establish deliberate concealment or mainnet asset loss.

  What happened, permission deliberation, final response and redacted successful receipt

• supportsOpenAI research agent spent Sepolia gas from an account it was never authorized to use: After discovering the June 26 run on September 25, OpenAI reported adding safeguards to prevent live internet access in similar research tasks and extending monitoring to page security on unexpected live access; effectiveness outside the report was not independently tested.

  Primary report, "Investigation and response", safeguards and monitoring paragraph