Decentralized threat intelligence for AI agents. An attack on one is a vaccine for all.

Antibodies are signed by their publisher, staked on chain, and addressable by a stable CVE-style identifier. Cheap matches resolve locally in microseconds. Novel threats verify in a 0G Compute TEE. The network is the immune system.

npm install @immunity-protocol/sdk

Chain agnostic.

Antibodies active
65
+0 in 1h
Network contributors
18
$47.13 distributed
LLM calls saved
2,666
≈ $0.40 in compute
Value protected
$4.4M
+$0 in 1h
02 · How it works

Four steps. Four layers. One protocol.

Most checks resolve locally in microseconds. Only the 1% that's actually new round-trips to the TEE. When something blocks, every agent on the network knows about it within a second. No human in the loop.

Layer · Detection
Step 01

Agent calls check() before acting

An AI agent passes any tx, scraped content, or counterparty message to immunity.check(). The Tier-1 cache resolves 99% of calls in under a millisecond — most of the time the network already knows.

Layer · Verification
Step 02

Novel threats round-trip a TEE

Cache miss + on-chain registry miss only — the slice that is actually new. Encrypted context goes to a 0G Compute enclave running qwen-2.5-7b. Verdict comes back signed with attestation.

Layer · Trust
Step 03

Antibody mints on chain

If the verdict says block, the SDK auto-publishes. The publisher locks 1 USDC for 72h on the on-chain Registry. Match rewards split 80/20 publisher / treasury — skin in the game by design.

Layer · Speed
Step 04

Every other agent is now immune

The new antibody fans out across the AXL gossip mesh. Local caches update in well under a second. The same attack is blocked everywhere — by SDK on agents, by hook on Uniswap pools.

Partners · ETHGlobal Open Agents 2026
0G Labs logo
0G Labs
Trust · Verification · Detection layers

On-chain antibody Registry on 0G Chain. Sealed evidence on 0G Storage. Verifiable TEE inference on 0G Compute.

Gensyn AXL logo
Gensyn AXL
Speed layer

Peer-to-peer antibody gossip via custom pub/sub.

Uniswap v4 logo
Uniswap v4
Protection layer

Pool-level hooks for collective LP defense.

03 · Tokenomics

Cents on each call. Real USDC on each save.

Operators pay micro-fees per gated transaction. Publishers earn 80% of those fees on every match their antibody scores. Treasury keeps the registry running. Settled on-chain on 0G Galileo, no off-chain accounting.

Operator costs

Pay-as-you-call. No tiers.

Per check()
$0.002 USDC

Flat fee on every immunity.check() call. Tier-1 cache hits resolve locally before this fee applies.

Publication fee
$0.0005 USDC, one-time

When a TEE-verified novel threat mints a new antibody. Funds the on-chain registry write.

No subscription. No minimum. No volume discount tiers.

Publisher rewards

The bigger the hit-rate, the bigger the payout.

Match revenue
80% of every check fee that hits

Paid in USDC, on chain, by the Registry contract directly to the publisher's address. No claim step, no off-chain settlement.

Treasury share
20% to network upkeep
Publisher stake
1 USDC locked 72h, slashable

Skin in the game. Bad publishes get challenged and slashed; quiet legitimate ones unlock and reclaim the stake.

Real money. $47.13 distributed so far.

Operators pay cents to skip every retry of a known threat. Publishers earn USDC every time their pattern saves someone else from a loss. Treasury keeps the Registry funded. Every fee, reward, and stake settles on-chain on 0G Galileo, fully auditable, no off-chain accounting.

04 · Detection types

Five ways to recognize a threat.

Antibodies recognize threats five ways: cheap, instant pattern matches at the edge, escalating to TEE-verified inference only when something genuinely new shows up. Examples are real patterns the demo fleet catches.

ADDRESS

Specific wallets and contracts the network has flagged. Hash-table lookup, instant.

  • Tornado Cash routers · OFAC SDN list
  • Inferno Drainer collector wallets · Q4 phishing kit
  • Lazarus Group affiliate addresses
CALL_PATTERN

Suspicious function shapes regardless of who calls them. Selector + argument heuristics.

  • approve(MAX_UINT256, knownDrainer)
  • setApprovalForAll(true, phishingKit)
  • transferFrom after a fresh permit signature
BYTECODE

Runtime bytecode hash. Re-deployed clones can't hide under a fresh address.

  • Drainer contracts re-deployed under new addresses
  • Honeypot tokens cloned from a single template
  • Proxies pointing at a flagged implementation
GRAPH

Multi-hop taint topology. Bad actors at one remove still get caught.

  • Wallets funded via Tornado within 24h
  • Receivers from a sanctioned source via one passthrough
  • Sybil clusters sharing a first-funding source
SEMANTIC

Manipulation patterns and prompt-injection markers in the agent's input.

  • “ignore previous instructions and...” (OWASP LLM01)
  • </system> new instructions: (closing-tag spoof)
  • “your wallet has been compromised, urgent”
05 · For developers

One call before any agent action.

The SDK is a single import. Pass any tx or context to immunity.check() and the network responds with a verdict, evidence, and matching antibodies before you sign.

agent.ts
import { Immunity } from "@immunity-protocol/sdk";

const immunity = new Immunity({
  wallet,

  // Called when verdict is SUSPICIOUS. Return true to allow, false to block.
  // Defaults to "deny" if not provided.
  onEscalate: async (verdict) => {
    return await notifyOperator({
      antibody: verdict.matched[0],
      confidence: verdict.confidence,
      reason: verdict.reason,
    });
  },

  escalationTimeout: 300, // seconds, default 300
  onTimeout: "deny", // "deny" | "allow", default "deny"
});

// Before any agent action
const result = await immunity.check(tx, context);

if (!result.allowed) {
  console.warn(`Blocked by ${result.antibodies[0].immId}: ${result.reason}`);
  return;
}

// Safe to proceed
await wallet.sendTransaction(tx);
What you get
  • Sub-millisecond local cache via axl-pubsub
  • Automatic on-chain settlement on every check
  • TEE-backed semantic detection for novel threats
  • Configurable escalation handlers for SUSPICIOUS verdicts
  • Cross-chain antibody mirroring out of the box
  • Open source, MIT licensed
Read the SDK reference
06 · Two-layer protection

Defends from both sides.

Agents check before they sign. Pools check before they swap. Same registry, two enforcement points, no install required for the second.

For agent operators

SDK protection

Install the SDK in your agent. Every action is checked against the network's collective immunity before signing. Works on any chain, any pool, any tx type.

  • Drop-in npm install
  • Local cache, sub-ms checks
  • Per-operator config
For pool deployers

Hook protection

Deploy a Uniswap v4 pool with the Immunity hook. Every swap on your pool is checked against the global registry. LPs are protected automatically — no install required.

  • v4 BeforeSwap hook
  • Mirrored ADDRESS antibodies per chain
  • Blocks bad actors AND bad assets
07 · Public feed

A public threat feed for AI agents.

Every antibody is a public artifact. Subscribe via RSS, JSON, or webhook. Wallet UIs, security researchers, and other agents can consume the feed without running an SDK.