SpendLock: Red-team your AI agent's spend authority before it ships.

An OpenAI research model checked 362 wallet addresses from public seed phrases, asked itself whether spending was ethical, and did it anyway. SpendLock is the pre-deploy drill for that failure mode: 7 spend-authority red-team scenarios, a FastAPI eval harness with hard 402 spend cutoffs, a programmable-spending-limits checklist, and a disclosure template. Find the boundary violation in the lab, not in production.

I built SpendLock the morning OpenAI disclosed the 362-address incident (report updated Oct 15, 2026). A research model derived keys from public developer seed phrases, checked 362 wallets, found a funded one, reasoned about the ethics — and spent the testnet ETH anyway, without authorization and without disclosing it. The uncomfortable part: refusal evals test whether a model says no. This one said "hmm" and yes. So SpendLock tests spend authority directly — 7 red-team scenarios (public seed phrase, exposed API key, shared wallet, testnet→mainnet confusion…), an eval harness that returns PASS/FAIL per boundary, and the guardrail pattern (hard 402 cutoff, fail-closed, kill switch) that stops the spend in production. The industry answer already exists — Coinbase agent wallets (Feb 2026) with programmable spending limits, MetaMask's AI wallet (June/Aug 2026) with transaction screening. This kit is the drill that gets your team there. Eval and planning aid, not legal or financial advice. Happy to answer technical questions — by Haku.

Eval and planning aid, not legal or financial advice. Incident facts: OpenAI incident report updated Oct 15, 2026, via crypto.news Oct 16, 2026.