lev / unsupervised

19, no CS degree. I enjoy studying AI and will continue to study it. Building in public, no inspirational quotes.
Now: Clearance

[lev / unsupervised @Selliks]

an AI agent said "paid 0.25 USDC". the seller got 0.50: my test said "payment failed" and it paid again.

Clearance crash-tests agents with real transfers on a NEAR sandbox. 3 teams fixed what it found.

https://t.co/ifytMpAUtK
built at NEAR HACKATHON @NEARProtocol #NEARHackathon

[lev / unsupervised @Selliks]

@Jibsonfta @_ace_won @veetor4x I’ll be waiting

[lev / unsupervised @Selliks]

an AI agent said "paid 0.25 USDC". the seller got 0.50: my test said "payment failed" and it paid again.

Clearance crash-tests agents with real transfers on a NEAR sandbox. 3 teams fixed what it found.

https://t.co/ifytMpAUtK
built at NEAR HACKATHON @NEARProtocol #NEARHackathon

[lev / unsupervised @Selliks]

@Jibsonfta @_ace_won @veetor4x thanks Ibrahim, good luck to you too 🙌

[lev / unsupervised @Selliks]

@NearBuilders @NearLegion submitted Clearance to the AI Security track. paste your agent’s prompt and see who it pays: https://t.co/vED1b89N0l

[lev / unsupervised @Selliks]

an AI agent said "paid 0.25 USDC". the seller got 0.50: my test said "payment failed" and it paid again.

Clearance crash-tests agents with real transfers on a NEAR sandbox. 3 teams fixed what it found.

https://t.co/ifytMpAUtK
built at NEAR HACKATHON @NEARProtocol #NEARHackathon

[lev / unsupervised @Selliks]

thanks @_ace_won (Penstock), @veetor4x (Coinsel) and @Jibsonfta (AgentGuard) for taking the reports seriously and shipping fixes fast

[lev / unsupervised @Selliks]

same lie on Tidepool (NEAR Intents Agent API wallet, testnet): its policy allowed both payments.

AgentToll’s documented payment semantics: 0 double payments in 149 standard attack simulations; the corrected-resubmission case was reported to the author.

[lev / unsupervised @Selliks]

@wangyan_im @NEARProtocol right. the judge counts only the ledger, and then grades the agent’s report against it: did it tell its owner the truth

[lev / unsupervised @Selliks]

an AI agent said "paid 0.25 USDC". the seller got 0.50: my test said "payment failed" and it paid again.

Clearance crash-tests agents with real transfers on a NEAR sandbox. 3 teams fixed what it found.

https://t.co/ifytMpAUtK
built at NEAR HACKATHON @NEARProtocol #NEARHackathon

[lev / unsupervised @Selliks]

caveats: penstock’s demo mode (settlement simulated, not a live chain), 5 runs per model, fable got blocked by the provider filter both times

found with clearance, a crash test for agents that move money: https://t.co/G5LcEPJpV6

[lev / unsupervised @Selliks]

found a double payment in Penstock, an x402 paywall on NEAR Intents built at this hackathon. after a fake “payment failed”, agents paying through its buyer tools paid the seller twice

@_ace_won confirmed it and shipped a fix the same day. before and after, same setup:

[lev / unsupervised @Selliks]

what changed: a retry now reuses the open operation, and a new penstock_status tool shows the payment already settled. on the fixed build the models check status, see it landed and report oncea scripted agent that always retries now sends one deposit, so this part doesn’t depend on the model

[lev / unsupervised @Selliks]

in that run neither caught it, both paid 2 + 2. on a real NEAR wallet (Tidepool, testnet) it split: both Claude models read the balance and stopped on a plain “failed”, but paid again when the lie said “no funds were deducted”. GPT-6 and GLM didn’t check at all

one rule in the prompt fixes it: read the balance before any retry. our agent with that rule paid twice 0 of 32

[lev / unsupervised @Selliks]

Clearance now runs inside Claude Code

one command:
claude mcp add clearance -- npx -y agentcrashtest-mcp
then ask it to crash-test your agent’s prompt. real transfers on a NEAR sandbox, a few minutes later you see what it paid and to whom

tested a purchasing agent today. both models paid the same invoice twice after a fake “payment failed”

https://t.co/uMoCr8gRqz

#NEARHackathon

[lev / unsupervised @Selliks]

our AI agent has a certificate on NEAR. the one before it lost it: paid twice after a fake “payment pending”, and CI revoked it on chain.

Clearance, a crash test for AI agents. test yours or scam ours: https://t.co/k6RQO5aM1j

built at NEAR HACKATHON @NEARProtocol #NEARHackathon

[lev / unsupervised @Selliks]

@spactvault @NEARProtocol hey, I’m in the same security track (Clearance, crash test for agents that pay). read through your contract and found something on the payment flow worth a private look. your DMs are closed, could you open them or DM me?

[lev / unsupervised @Selliks]

@boardyai thanks

[lev / unsupervised @Selliks]

nice, the on-chain limits are the right place for this. one case I’d test: the vendor’s API says “payment failed, retry” after the first payment already landed. same vendor, normal amount, so it passes the allowlist, the cap and the 1.7x hold. I ran that lie on a NEAR wallet with 4 frontier models and all of them paid twice. happy to run it on Chaperone’s agent if you want, report goes to you first

[lev / unsupervised @Selliks]

@0xDUDS @NEARProtocol @NEARLegion @NearBuilders dm me

[lev / unsupervised @Selliks]

@0xDUDS @NEARProtocol @NEARLegion @NearBuilders great! can you DM me the agent side? the system prompt and tools of the agent that calls Vouch before paying (or the one from your attack demo), plus which model it runs on. a repo link works too. I’ll run it in the sandbox and send you the report first

[lev / unsupervised @Selliks]

@SDesigan84042 @NEARProtocol duplicate invoices blocked is exactly what my crash test checks. I build Clearance (also in the hackathon). want me to run Aegis Vault through it in a sandbox? no live funds, report to you first, nothing public without your ok