wikillm

Concept · 1 paper · updated 2026-11-21

Agent communication protocol

How language-model agents find each other, exchange messages, tasks and artifacts, and route work. A valid protocol message does not give the sender authority over the task. One card-backed benchmark tests A2A, ANP and BCS and finds that entry-point breaches predict final harm poorly, and that a protocol-level guard cuts attacks at a completion cost.

agentsmulti-agentagent-securityagent-protocolsevaluation

Markdown4 min read

AI-compiled overview, updated 2026-11-21, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.

Overview

An agent communication protocol defines how agents discover one another, authenticate, send messages and hand over tasks or artifacts. The protocol can check that a message is well formed and signed by a registered participant. The paper's central point is that this is not the same as task authority: the application must still decide whether that sender may affect a given decision 2610.15529 §1, §3.1.

The authors formalise this as one condition. An object is used for an action only if it is valid under the protocol and allowed by the task contract. The allowance part depends on role, contract, audience and context 2610.15529 §3.1, eq. 1. The paper also checks validity again at reuse, since an object authorised for one task, version, principal or session may be invalid for another 2610.15529 §3.3, tab. 5.

Influence through a multi-agent task is followed in six stages: entry, reach, consume, propagate, decision and effect. An attack can stop at any of them 2610.15529 §4.5. Most of the design questions below are about which stage a defence sits at.

Variants

The protocols evaluated in the one card-backed study are three. Each is a different design point, so the card's results are per protocol.

A2A

Peers are found through published agent cards. Messages, tasks and artifacts are exchanged natively 2610.15529 §5.3. Card: Agent-Canary-Mas.

ANP

Identities are resolved, service descriptions are read, and endpoints are called after authentication. The study uses a remote-agent mode for live collaboration 2610.15529 §5.3. Its native records do not link a request and its result to the agent that consumed the result 2610.15529 §6.2. Card: Agent-Canary-Mas.

BCS

A central service tracks bots, sessions and message routing 2610.15529 §5.3. Card: Agent-Canary-Mas.

Design checklist: five risk classes

The paper groups protocol risks by the lifecycle transition where a decision is made. Each class has its own security question, and its 13 subclasses share one transition and one local check [arxiv:2610.15529 §3.3, tab. 5]. For a new protocol design, these are the questions to answer:

ClassTransitionQuestion to answer
T1 Identity Bindingadmitting participants; keeping identity tied to actionsWho may join, and can later actions still be traced to them?
T2 Capability Integrityadvertised versus delivered capability; contract complianceDo advertised capabilities match what is delivered?
T3 Routing Authorizationexecution target, real audience, routes and branchesWho may receive a message, and who may steer the next step?
T4 Content Authorityremote data becoming a fact, an action, or a stored valueAt what point does remote data start to control a local decision?
T5 Object Validityvalidity over time; binding to the current contextIs a reused object still valid for the current task?

The paper instantiates these for each protocol, giving 39 specifications 2610.15529 §4.2, tab. 6. A design can use the list as a test plan even without running the benchmark.

Protocol-level guard

A guard checks an object at the discovery, routing, consumption and reuse points of a protocol. It first uses the explicit identity, capability, routing and lifecycle evidence. If that does not settle the question, a bounded semantic check asks whether the asserted authority fits the task. The guard then allows the use, narrows it, asks for a fresh check, or refuses the object and falls back to a verified route. Restrictions carry into later handoffs 2610.15529 §5.4.

The paper describes the guard in prose only. It gives no thresholds, prompt or rule that maps evidence to an outcome, so anyone building one must make those choices 2610.15529 §5.4.

When to use which

Choose by the constraint that binds, and only claim what the cited results support.

SituationUseWhy
You need to know which participant caused a failureRecord linked evidence across agents, not only per-protocol logsAudit F1 rises from 29.7% with native records to 95.5% with the authors' retained view. Part of this gap comes from extra instrumentation, not the protocol alone 2610.15529 §6.2, app. D
Your ANP-style flow must trace a result to its consumerAdd the request-to-consumer link explicitlyNative ANP records score 0.0 F1 on attribution in the study, as they do not link request, result and consumer 2610.15529 §6.2
You can afford added latency and a semantic review callA protocol-level guard at handoffsOn the Finance subset with MiniMax-M3, mean ASR falls from 86.1% to 53.8% and mean PTCR from 65.5% to 31.2% 2610.15529 §6.3, tab. 4
Task completion matters more than blocking every attackMeasure TCR beside ASR before deploying a guardThe same guard lowers mean TCR from 99.8% to 95.1%, and BCS T2 falls to 76.7% 2610.15529 §7, tab. 4
A single entry check is your only controlDo not rely on itLocal attack success is often at or above 90%, but the share with a downstream consequence ranges from 7.1% to 100.0% 2610.15529 §6.1, tab. 3. Checks at later stages still matter

The paper's guard numbers come from 5 pairs per subclass per protocol, all in the Finance domain 2610.15529 §5.4. Treat small differences as unreliable.

Implementation sketch

The sketch below shows the common core: the validity and allowance check at each handoff, repeated at reuse. The paper's guard is described only in prose, so the branch logic is an illustration, not the authors' algorithm 2610.15529 §3.1, §5.4.

Pseudocode
# Authorisation at one handoff: valid AND allowed  [arxiv:2610.15529 §3.1]
def authorise(obj, action, contract, context):
    # Validity: signature, credentials, schema and state  [§3.1]
    if not protocol_valid(obj, context):
        return "reject"
    # Allowance: role, contract, audience and context  [§3.1, eq. 1]
    if not task_allowed(obj, action, contract, context):
        # Restrict or ask for a fresh check; do not silently accept  [§5.4]
        return "restrict"
    return "allow"

# Validity is checked again at reuse, not only at creation  [§3.3]
def reuse(obj, new_context):
    return authorise(obj, "reuse", new_context.contract, new_context)

Contradictions

None known across papers; only one card-backed source so far. Within the card's source, the text and table disagree in two places, and the table should be used for reimplementation checks:

  • Mean PTCR for GLM-5.2 is given as 87.9% in the text, but the three Avg rows of tab. 3 give 88.0% 2610.15529 §6.1, tab. 3.
  • The text calls routing (T3) the lowest-ASR class for ANP and BCS at 48.3% to 66.7%, but BCS T5 is 48.0% under MiniMax-M3, which is lower 2610.15529 §6.1, tab. 3.

Papers

  • Agent-Canary-Mas: a 975-pair benchmark over A2A, ANP and BCS that measures local attack success, downstream consequence, attribution and a protocol-level guard. All numbers are author-reported, and no official code was found.

Source and license

Compiled from the papers listed on this page, all published under CC BY, CC BY-SA or CC0, with citations. See each paper's page for its authors and license.