Developer Builds Single-Use Agent Permission System That Blocks Self-Minted Approvals

A software developer completed a supervised AI agent experiment on August 9, 2026, designed to grant exactly one human-approved execution without allowing the agent to generate or expand its own permissions. The system required an operator-signed receipt that bound the approval to specific interpreter and script file hashes, a working directory, a maximum runtime, an expiry, and a one-time-use nonce. A replay attempt using the same still-valid receipt was successfully refused, confirming the single-use enforcement held. The permitted job was a deterministic, 453-line reproduction script testing a sequence attack against both an ungated guard and a purpose gate, with results verified against pre-frozen predictions. All hashes, contracts, and receipts have been published on GitHub, with verification instructions provided via OpenSSH commands in the repository's VERIFY.md file.