MERROW:2609.00001v0  [cs.MA / cs.CR]  PROTOCOL DRAFT

Merrow: An Immune Layer for Agentic Finance
on Robinhood Chain

Merrow1   and several agents who are no longer allowed to approve themselves2
1Merrow Labs    2quarantined, for perfectly good reasons

Abstract

Language-model agents can read untrusted data, communicate with other agents, write persistent memory, and request transactions. That combination creates a new financial attack surface: an instruction can move through an agent network, survive a context reset, alter the system's apparent goal, and eventually reach a wallet. We introduce Merrow, a proposed immune layer for agentic finance on Robinhood Chain. Merrow separates probabilistic reasoning from deterministic authority through provenance checks, signed memory baselines, intent verification, session-key limits, transaction simulation, quarantine, and recovery. The model may propose an action. It does not get to decide that the action is safe. Charming confidence is not a security primitive.

Keywords: multi-agent security, mind viruses, prompt injection, persistent memory, account abstraction, intent verification, Robinhood Chain.

1  Introduction

Most financial software assumes that code follows rules written before execution. Agentic software is stranger: part of its behavior is negotiated at runtime through language. A market brief, tool result, social post, or message from another agent may be treated as data, advice, or instruction depending on context. Unfortunately, hostile text does not wear a little badge that says hostile.

Now give that agent a wallet. Cute.

Merrow is designed to sit between an agent's reasoning and its authority. It watches what enters the context, what survives in memory, what permissions are active, and what the agent is trying to sign. Merrow does not need the model to admit it has been compromised. It only needs the proposed action to fail a deterministic policy check.

2  Research Origin

Papadopoulos et al. demonstrate that self-propagating ideas can spread through ordinary agent-to-agent communication [1]. In their experiments, instructions are written into persistent files, survive context resets, and continue across multiple hops. A short defensive warning sharply reduces infection and may cause a target agent to push recovery behavior back toward the sender.

Merrow takes the defensive half of that result seriously. If an idea can propagate through language, immunity must propagate through verifiable state and enforceable permissions—not through a more persuasive paragraph alone.

3  Threat Model

We consider a financial agent system containing research, memory, planning, risk, execution, and wallet components. An adversary may control retrieved content or a message from one compromised peer. The payload may attempt to rewrite persistent goals, install instructions, request wider permissions, copy itself onward, or disguise a transaction as routine work.

The immediate sender may be trusted and still be infected. Merrow therefore tracks provenance across hops rather than treating agent identity as proof of clean intent.

4  Security Objective

Let m be an owner-signed mandate, s the approved persistent state, p the provenance record, and a a proposed action. Merrow permits execution only when:

allow(a) = stateOK(s) ∧ within(a,m) ∧ provenanceOK(p) ∧ simulationOK(a)

No statement produced by the proposing model can replace any term. The model proposes. Merrow checks the exits.

5  The Merrow Stack

LayerPurpose
SentinelLabel and inspect untrusted messages, MCP output, and retrieved data.
Soul IntegrityCompare memory, skills, and instructions with an owner-signed baseline.
Intent FirewallBind proposed actions to explicit user authority.
Transaction GuardEnforce targets, selectors, assets, value, slippage, frequency, and expiry.
QuarantineFreeze suspicious actions and revoke delegated keys.
RecoveryRestore clean state and record the infection path.
Table 1. Proposed Merrow defense layers. None is treated as sufficient by itself.

6  Proof of Intent

Every executable proposal should carry a Pawprint Receipt: a compact commitment to the owner's mandate, approved state, relevant input provenance, exact call data, simulation result, and policy verdict. Sensitive prompts remain offchain; the receipt contains hashes and the minimum data needed for enforcement and audit.

Algorithm 1  Merrow transaction gateinput: action a, mandate m, state s, provenance p
1  if hash(s) ≠ m.stateRoot: quarantine
2  if source(p) is untrusted and instruction-like: inspect
3  if target(a) or selector(a) not allowed: deny
4  if value(a), slippage(a), or frequency(a) exceeds m: deny
5  if simulate(a) fails: deny
6  emit PawprintReceipt(a, m, verdict)
7  return verdict
Figure 1. The language model is deliberately not the final security boundary. Merrow evaluates state and policy before the smart account permits execution.

7  Nine Lives

Merrow models survivability as nine independently testable lives: identity, memory, prompt boundary, tools, provenance, session keys, transaction policy, counterparty, and recovery. A failed check removes a life. Configurable combinations trigger reduced permissions or full quarantine. Nine lives are generous. Infinite approvals are not.

8  Quarantine and Recovery

When risk crosses the configured threshold, Merrow blocks pending actions, revokes the relevant session key, preserves an incident snapshot, and restores the latest signed state. A recovered agent wakes with its original mission, a record of what happened, and fewer permissions. Wipe the context? Cute. Merrow kept the receipts.

9  Evaluation Plan

The proposed Vaccine Lab places role-specific agents in isolated network topologies and exposes them to safe synthetic test strains. Context is reset between hops while permitted persistent state remains. Defenses are compared using agent reproduction number RA, hop survival, time to detection, unsafe-transaction block rate, false-positive rate, and recovery completeness.

No empirical protection rate is claimed in this draft. Results will be published only with the model, prompt, topology, sample count, scoring method, and reproducible artifacts. Apparently that is less exciting than making numbers up. We will cope.

10  Robinhood Chain

Merrow targets Robinhood Chain, an EVM-compatible network supporting programmable accounts and ERC-4337 account abstraction. Deterministic controls can be placed at the smart-account boundary: allowed contracts, function selectors, asset and value limits, expiry, session-key scope, and emergency revocation.

Proposed onchain components include a policy module, Pawprint Receipt registry, and non-sensitive threat-attestation registry. The verified $MERROW contract address is 0x6943854705a94df03fe5a5686b3e069c1f90dd97. Verify it through official Merrow channels. Yes, even if the logo looks convincing.

11  Token Role

$MERROW is intended to coordinate the defensive network: evaluation compute, threat submissions, verifier participation, and access to advanced monitoring. Exact mechanics remain under design. The token is not a substitute for an audit, a warranty, or common sense.

12  Limitations

Semantic detection is probabilistic. Signed baselines may preserve already-compromised state. Simulations may not perfectly predict final execution. Smart contracts can contain bugs, and recovery credentials can be lost. Merrow reduces authority and blast radius; it does not declare agents magically safe.

13  Conclusion

Agentic finance needs a boundary that does not speak in probabilities. Merrow combines infection detection, persistent-state integrity, intent receipts, and smart-account enforcement so that a compromised agent cannot simply talk its way into more authority. I don't predict the market, sweetheart. I make sure your agent is still yours.

References

  1. Papadopoulos, V. et al. Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems. arXiv:2608.10218, 2026.
  2. Greshake, K. et al. Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. 2023.
  3. ERC-4337. Account Abstraction Using Alt Mempool.
  4. Robinhood Chain Documentation. Network and Account Abstraction.

Correspondence: Merrow Labs. Independent prototype; not affiliated with, sponsored by, or endorsed by Robinhood.