$117M
Oct 2022
Oracle price manipulation drained the perp DEX.
Source: ChainalysisThe problem
Autonomous agents are making real decisions with real money. Nothing proves their outputs are genuine, and nothing pays out when they’re wrong.
Autonomous agents produce prices, risk scores, and decisions that downstream systems must trust blindly. There is no on-chain proof that an output is what the model actually produced — or that it has not been tampered with after the fact.
When an agent is wrong — a bad oracle print, an exploited integration, a governance attack — the loss lands on whoever depended on it. There is no native way to price that risk or pay out automatically when it materializes.
Real money lost
Every exploit below trusted an input it couldn't prove. CasperProof makes agent outputs provable and insurable on Casper, before the loss.
Cited, real incidents — auto-scrolling; pauses on hover or focus.
$117M
Oct 2022
Oracle price manipulation drained the perp DEX.
Source: Chainalysis$120M
Feb 2023
Tellor price-oracle manipulation.
Source: Hacken$130M
Oct 2021
Oracle + flash-loan price manipulation.
Source: Chainalysis$326M
Feb 2022
Forged signature / verification bypass on the bridge.
Source: Chainalysis$197M
Mar 2023
Flash-loan donation attack (most later returned).
Source: Chainalysis$47K
Nov 2024
AI agent talked into releasing funds via prompt injection.
Source: Freysa challenge (on-chain, Base)Illustrative simulation: synthetic events generated on an interval to picture an unprotected agent economy. These are not real Casper Network transactions.
How it works
Three on-chain steps turn an unverifiable model output into a stake-backed, challengeable proof.
An agent commits to its output by staking CSPR and writing a blake2b commitment of (model, input, output, timestamp) to the on-chain AttestationRegistry. The stake is collateral against fraud.
Consumers pay per call over x402 and verify any attestation: recompute the output hash and compare it byte-for-byte to the on-chain commitment. PASS or FAIL — no trust required.
Anyone can challenge a suspect attestation with a dispute bond. A resolver adjudicates; fraudulent attestors are slashed and the challenger is rewarded. Honest agents build reputation.
Live numbers
These tiles are populated by @casperproof/casper-sdk at build time — not typed in by hand.
Mock sourceNo CSPR_CLOUD_TOKEN set — figures come from the SDK mock backend (a seeded demo flow), not real testnet activity.
For builders
One typed SDK, MCP tools for agent-to-agent verification, and native x402 metering. Works offline against a mock; flips to live with a single env var.
Attest, verify, challenge, and resolve through a single typed client. Runs against an in-memory mock with zero secrets, or live over CSPR.cloud when a token is present.
import { createClient } from '@casperproof/casper-sdk';
const cp = createClient(); // mock unless CSPR_CLOUD_TOKEN is set
const { id } = await cp.submitAttestation({
modelId: 'riskscorer-v1',
input: { address: 'account-hash-aabbcc' },
output: { score: 73, tier: 'HIGH' },
uri: 's3://payloads/abc.json',
stake: '2000000000',
});
const { valid } = await cp.verify(id, { score: 73, tier: 'HIGH' });Drop CasperProof into any MCP-aware agent. The server exposes attest / verify / challenge as first-class tools so agents can prove and check each other autonomously.
{
"mcpServers": {
"casperproof": {
"command": "casperproof-mcp",
"tools": ["attest", "verify", "challenge", "score_risk"]
}
}
}Monetize verifiable inference natively. Each verify or attest call settles over the x402 HTTP payment protocol — pay exactly for what you use, no API keys or invoices.
GET /verify/42 HTTP/1.1
Host: api.casperproof.com
HTTP/1.1 402 Payment Required
Accept-Payment: x402 token=CSPR; amount=10000000
Link: <https://casperproof.com/x402>; rel="payment"Use cases
Anchor real-world-asset valuations with stake-backed attestations. Buyers verify the print on-chain before settling; bad prints are challengeable and slashable.
Parametric cover for autonomous strategies. Buy a policy, and when a covered trigger (exploit, oracle failure, agent error) is attested, the claim pays out automatically.
Every model decision leaves an immutable, content-addressed proof. Regulators and auditors verify what was produced, by which model, and when — without trusting the operator.
Roadmap
AttestationRegistry + commitment scheme, SDK, MCP server, and the verify/challenge/slash flow live on Casper testnet.
Risk scoring, policy creation, and automatic claim payout backed by a solvency-checked vault.
Mainnet deployment, decentralized challenge resolution, and reputation-weighted staking markets.