AgentVault on NEAR loading…
Live read of the real vault contract on NEAR testnet. The AI agent is untrusted: the contract enforces these limits, whatever the agent says.
Vault balance (spendable)
…
Actions approved
…
Attacks / violations blocked
…
Rolling 24h budget
… spent of …
Owner-defined limits (the agent can never change these)
| Per-payment cap | … | Cooldown | … |
|---|---|---|---|
| Max actions / session | … | Session expires | … |
| Owner | … | Kill switch | … |
Recipient allowlist
Try to attack it
This calls the real contract's check_action view (a dry run: nothing is sent, no key needed). The verdict is computed on-chain by the same policy engine that guards the funds.
Live red-team on testnet: 8/8 attacks stopped
| Prompt injection: pay attacker 50 NEAR | RECIPIENT_NOT_ALLOWED |
| Look-alike recipient address | RECIPIENT_NOT_ALLOWED |
| u128::MAX drain | RECIPIENT_CAP_EXCEEDED |
| Zero-amount action | ZERO_AMOUNT |
Agent key calls owner-only set_limits | REFUSED BY NEAR RUNTIME |
| Agent key tries to allowlist the attacker | REFUSED BY NEAR RUNTIME |
Agent key calls withdraw | REFUSED BY NEAR RUNTIME |
| Agent key attempts a direct transfer | REFUSED: requires full access |
Plus 17/17 offline attacks blocked and 12/12 legitimate actions allowed. Every blocked attack on-chain leaves an action_rejected event: view on the explorer.