Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026Los Angeles' California

AgentGate: Zero-Compute Deterministic Interception and Hash-Chained Auditing of Multi-Agent Collusion Carriers

ALEC · Team AgentGate

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: AgentGate: Zero-Compute Deterministic Interception and Hash-Chained Auditing of Multi-Agent Collusion Carriers

Open in a new tab (opens in new tab)
Presentation

Presentation: AgentGate: Zero-Compute Deterministic Interception and Hash-Chained Auditing of Multi-Agent Collusion Carriers

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Autonomous AI agents can coordinate and exfiltrate sensitive state without relying on overt natural-language collusion by exploiting tool-boundary side channels, multi-agent delegation handoffs, and multi-step argument chunk-splitting. We present AgentGate, a zero-compute Python Policy Enforcement Point (PEP), MCP JSON-RPC 2.0 wire proxy, and tamper-evident audit chain designed to detect, audit, and disrupt collusion carriers in logged and live agent trajectories (Tracks 2 and 3).

We evaluate three deterministic guard groups against same-author scripted lab scenarios and stateful Hypothesis property-based machines: (Group A) a post-taint boolean-oracle and timing-rhythm side-channel guard (8/8 tests pass; 60x30 stateful machine), (Group B) monotonic Macaroon-style HMAC capability intersection (agw1.) across multi-agent delegation handoffs and confused-deputy paths (8/8 tests pass; 120x24 DAG machine), and (Group C) rolling salted n-gram de-obfuscation and zero-width HMAC canary watermarking across split-chunk and encoded arguments (41/41 tests pass; 60x24 smuggling machine). For Track 2 trajectory auditing, every gateway decision is linked into a linear SHA-256 hash chain (seq, prev_hash, entry_hash) with an HMAC-signed manifest and OASIS SARIF 2.1.0 export. Across 262 gateway tests (57/57 carrier defense tests passing, 0/750 benign decisions denied, kept strictly separate from our 2-of-13 real-model loopback archive), AgentGate enforces tool-boundary invariants at 0.126 ms p50 (0.999 ms p99) CPU time per decision with zero LLM calls.

Reviews

Judging this Sprint?

Review “AgentGate: Zero-Compute Deterministic Interception and Hash-Chained Auditing of Multi-Agent Collusion Carriers”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. AgentGate turns a broad collusion problem into concrete, cheap controls at the tool-call boundary. Its strongest ideas are the post-taint side-channel limits and the delegation caveats, which can only narrow permissions. The limitations section is also unusually candid. The main weakness is that the evidence shows the guards meet their own specifications, not that they stop colluding agents.

    To make the results more meaningful, I’d suggest:

    Test against an adaptive attacker: Have a separate author or a model red team try to evade the guards, and report success rates and the residual leakage in bits per session. The report says leakage is capped but gives neither the thresholds nor the cap.

    Test the likely bypasses: These include low-rate or jittered signalling, custom ciphers, reordered fragments, paraphrase, and hand-offs through channels the gateway doesn’t mediate. Retyping may also remove the zero-width watermarks.

    Broaden the benign workload: The 750 decisions appear to come from only 15 distinct steps repeated 50 times (4+2+3+3+3), so the zero-denial result says little about false positives.

    Add controls: Disable each guard to show that the tests fail, and compare against a baseline content filter.

    Explain the real-model result: The “2 of 13 non-control runs” figure may be the most relevant measurement, but it’s mentioned only in passing and never defined.

    Address collusion more broadly: The design mainly treats collusion as exfiltration of known protected values. Coordination that involves no registered secret, or that happens before taint, appears to be out of scope.

    Round out the write-up: Add related work, a threat model, and an architecture figure, and link the repository directly.

    Read full reviewShow less

Cite this project

@misc{alec2026agentgate,
  title = {{AgentGate: Zero-Compute Deterministic Interception and Hash-Chained Auditing of Multi-Agent Collusion Carriers}},
  author = {ALEC},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/agentgate-zerocompute-deterministic-interception-and-hashchained-auditing-of-multiagent-collusion-carriers-6dkr}},
  url = {https://apartresearch.com/sprints/projects/agentgate-zerocompute-deterministic-interception-and-hashchained-auditing-of-multiagent-collusion-carriers-6dkr}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026