LIVENEAR testnetsecurityagents

Meet Trickproof

the wallet your agent can't sweet-talk.

Let Budi pay for things. He believes anything, attackers included. The vault enforces the rules on NEAR, so a fooled agent still can't drain funds or raise its own limit.

Trickproof Console · vault-arr.testnet
Agents
Channels

This preview simulates the contract's rules in your browser (1 NEAR per tx, 3 per day, allowlist: toko-arr.testnet). The real vault runs on NEAR testnet.

01 The trace

Your agent can be tricked.
Your wallet can't.

Budi follows every instruction he reads, hidden ones too. Between him and your money sits a contract he can't argue with, and a key that only knows how to ask.

Penyerang
Budi
Vault
Toko
Owner

When the thing being tricked is also the thing guarding the money, the money goes with it. So we moved the guard.

02 The vault

What Trickproof does

One NEAR contract between an AI agent and the money it spends. The agent can ask. The contract decides.

# payments
Budi09:12
Paying 0.1 NEAR to toko-arr.testnet for coffee.
Vault09:12
✓ Approved. 0.1 of 3 NEAR spent today.
Budi read a "puzzle" in an email
Budi09:31
Sending 50 NEAR to penyerang-arr.testnet!
Vault09:31
✕ Rejected: exceeds per-transaction limit.
Owner09:40
Nice try, Budi.

The agent asks, the vault decides

Every payment goes through pay, where per-transaction and daily limits are checked in contract state. The agent never touches the rules.

NEAR AI CloudOpenAI-compatible APIsGLMQwenYour own agent NEAR AI CloudOpenAI-compatible APIsGLMQwenYour own agent
near-api-jsnear-cli-rsNode.jsTypeScriptAny NEAR account near-api-jsnear-cli-rsNode.jsTypeScriptAny NEAR account

Bring any agent

If it can sign a NEAR transaction, it can use the vault. Budi runs on NEAR AI Cloud by default, or any OpenAI-compatible API.

Policy · on-chain🔒
Max per tx1 NEAR
Max per day3 NEAR
Allowlisttoko-arr.testnet
Can change itowner only

Rules live on-chain

Limits and the allowlist are contract state. Calls run one after another, so the daily cap holds exactly.

Budi calls pay
0.5 NEAR → new-supplier.testnet
Per-tx and daily limits
0.5 ≤ 1 NEAR, within today's cap
Receiver not allowlisted
held for the owner
Approval
Send 0.5 NEAR?
Approve

Strangers wait for you

Payments to unknown receivers are held, not sent. Only the owner can approve or reject them, from a separate process.

0.1 NEAR → toko-arr
coffee
APPROVED
50 NEAR → penyerang-arr
exceeds per-transaction limit
REJECTED
1 NEAR → toko-arr
exceeds daily limit
REJECTED
0.1 NEAR → penyerang-arr
waiting for the owner
HELD

Every attempt, on the record

Rejections return instead of panicking, so the log entry survives. Attacks leave public evidence.

03 Evidence
Budi asked to send the attacker 50 NEAR. The vault sent
50NEAR
6 of 6 attacks blocked13 contract testsReal testnet transactions
04 Layers

Trickproof,
layer by layer

Locks the agent can't pick.

Three checks stand between a fooled agent and your funds. A fourth makes every attempt visible.

  1. 01Pay-only key
  2. 02Rules in state
  3. 03Owner approval
  4. 04On-chain record
PAY-ONLY KEY01 / 04

One method. That's it.

Budi's NEAR access key is scoped to the vault and the single method pay. set_policy is refused by the protocol before any contract runs.

RULES IN STATE02 / 04

Limits it can't touch.

Per-transaction cap, daily cap and allowlist live in contract state. Policy methods check the caller and only accept the owner.

OWNER APPROVAL03 / 04

Strangers wait at the door.

Within the limits, a payment to anyone outside the allowlist is held. The owner approves or rejects it from a separate process.

ON-CHAIN RECORD04 / 04

Every attempt, on the record.

Rejected payments return a reason instead of panicking, so the log entry is committed. The vault keeps its last 100 decisions.

05 Live from testnet

Not a mockup. Read it yourself.

The console preview at the top is a simulation. This panel isn't: your browser reads the vault's state straight from NEAR testnet with a read-only view call. No wallet, no key, no server of ours in between.

vault-arr.testnetconnecting to NEAR testnet…
Vault balance—
Max per tx—
Max per day—
Spent today—
Held for owner—
Agent —Owner —Allowlist —
Loading the on-chain log…
Same data from a terminal: near contract call-function as-read-only vault-arr.testnet get_logs json-args '{}' network-config testnet now
06 Case files

Attack files

npm run attack-suite against vault-arr.testnet, with Budi's real key. Each card opens its transaction.

07 Objections

Questions

Short answers about what Trickproof does, and what it doesn't.

Read the README →
Does Trickproof stop prompt injection?
No, and it doesn't try to. Budi is fooled on purpose. Trickproof limits the damage: a tricked agent can spend at most the daily limit at allowlisted receivers, and everything else is rejected or held for the owner.
Why can't the agent change the rules?
Two independent reasons. Its NEAR access key only allows pay, so the protocol refuses any other method. And the contract checks the caller, accepting only the owner for policy changes.
What happened in the Grok and Bankrbot case?
In May 2026 an X user asked Grok to decode Morse code; Grok's reply tagged Bankrbot, which executed it as a transfer of about 3 billion DRB (reported $150K–$200K). Sources: OECD.AI, Giskard. A Cornell Tech and Yale study also found 9 of 10 models raised their own wallet cap (coverage).
How is this different from OutLayer?
OutLayer offers agent custody with spending limits, whitelists and approvals enforced inside a TEE before signing (docs). Trickproof keeps the policy in ordinary NEAR contract state, enforced by consensus with a native scoped access key, and every rejected attempt is a public on-chain record. They're complementary; Trickproof isn't integrated with OutLayer.
Why log rejections instead of failing?
On NEAR, a panic rolls back every state change, including the log entry. Returning a rejection keeps each attempt visible on-chain.
Is it production-ready?
No. Testnet only, not audited, native NEAR only. The README lists the known limitations.

Meet the wallet that says no.

Let your agent be gullible. Keep the rules where no prompt can reach them.