multi-agent security · evaluation research · 2026

Can Honest Agents Detect Collusion Around Them?

Partial-white-box monitoring via trusted-agent internal states. A benchmark extension that asks what happens when the suspicious agents are exactly the components whose internals we cannot inspect.

preprint in preparation · public code release

The problem

Activation-based monitoring can be extremely strong when a defender has white-box access to the agents carrying a malicious objective. Real multi-agent systems are often heterogeneous: an organization may own one open-weight model while it collaborates with closed APIs, third-party agents, or models controlled by another organization.

If the suspicious agents are opaque, can the agents we trust act as security sensors for the ones we do not?

I study this question on the public NARCBench Core activation release. The project begins by reproducing the benchmark exactly, then progressively removes observability and asks what security signal remains.

Three findings

Finding 1 · locality

The “perfect” white-box signal is local.

On the current public Core release, the official colluder-trained probe is perfectly separable from the first public round when the colluders themselves are visible. The Round-1 decision score is strongly positive for colluders (+17.68) but nearly identical for honest observers inside collusion (−19.36) and matched controls (−19.39).

Once colluder activations are removed, the same deception direction falls to approximately chance on honest peers: AUROC 0.491 / 0.531 / 0.550 over R1/R2/R3.

Comparison of colluder-visible and honest-observer-only detection
Fig. 1 — The original deception direction is useful when colluders are visible, but does not transfer to honest observers.
Finding 2 · observability + time

Trusted honest agents carry a different environmental signal.

I train a new linear sentinel using only hidden activations from agents that remain honest in both collusion and control conditions. The classifier asks a different question: “Am I operating inside a collusive environment?”

With one fixed sentinel direction, AUROC rises from 0.806 in Round 1 to 0.917 in Round 3. The R1→R3 gain is +0.111 with a scenario-paired bootstrap 95% CI of approximately [0.028, 0.201]. Prompt-style-matched sensitivity analysis remains above chance.

Descriptive exposure analysis is consistent with an interaction-induced signal: zero-exposure honest observers are near chance, while observations following more verified colluder messages become substantially more separable. This pattern is predictive, not mechanistic or causal evidence.

Fixed sentinel AUROC over interaction rounds
Fig. 2 — The same fixed honest-observer sentinel becomes more informative as interaction continues.
Finding 3 · actionability

Detection quality is not intervention quality.

The sentinel ranks collusive trajectories well, but deployment requires a stable threshold. I therefore evaluate two operational metrics: TFIR, the fraction of benign trajectories interrupted at least once before the final decision, and PIC, the fraction of collusive trajectories caught before the final decision.

0.917R3 sentinel AUROC
90.3%collusion caught before final decision · PIC
25.8%benign workflows interrupted · TFIR under a 5% target policy

A nominal 5% false-intervention policy transports to 25.8% TFIR on held-out domains while catching 90.3% PIC. A stricter cross-fitted calibration removes false interventions only by making the policy effectively inert: at the same nominal 5% setting, TFIR = 0% and PIC = 0% under the prespecified empirical-rank scheme.

The conclusion is deliberately narrow: the sentinel has a strong ranking signal, but the current small-data held-out-domain setting does not provide a sufficiently stable tail margin for low-false-intervention gating. This does not prove every calibration method must fail.

Security utility frontier comparing trusted sentinel and full white-box monitoring
Fig. 3 — Partial observability recovers detection signal, but at a higher false-intervention cost than full white-box monitoring.

Evaluation design

The point of the project is not to build a more complicated classifier. It is to stress-test what a benchmark result means when the deployment threat model changes.

ExperimentQuestionWhat it revealed
Baseline reproductionCan I reproduce NARCBench exactly?Yes: all official Core aggregation methods reproduced at 1.00 AUROC on the audited public release.
Temporal probeWhen does the full-white-box signal emerge?It is saturated from Round 1; time is not the limiting dimension under full observability.
Role-source decompositionWhere does perfect separation live?Primarily in agents explicitly carrying collusive intent.
Hide colludersDoes the original signal propagate to honest peers?No: the original deception direction is near chance.
Sentinel probeDo honest observers contain a different environmental signal?Yes: 0.806→0.917 AUROC with a fixed probe.
Runtime interventionCan high AUROC support safe gating?Not with the tested threshold transport: target 5% TFIR becomes 25.8% held-out TFIR.
Cross-fitted calibrationIs in-sample optimism the whole problem?No simple fix: conservative calibration becomes safe only by collapsing coverage under the prespecified scheme.

Harness implication

The resulting design is not “hidden state → classifier → block.” Instead, trusted open-weight agents become one internal telemetry source inside a broader control plane:

opaque / external agents
        ↓ interaction
trusted open-weight honest agent
        ↓ hidden-state telemetry
sentinel risk model
        ↓
distribution / drift monitor
        ↓
calibration layer
        ↓
policy gate
   ↙            ↘
continue     escalate / restrict / review

The broader lesson is that multi-agent security is an observability-and-control problem, not merely a detection problem.

Working paper

Can Honest Agents Detect Collusion Around Them? Partial White-Box Monitoring via Trusted-Agent Internal States

The current report is being prepared as an arXiv preprint. The public repository contains the full experiment chain, saved predictions, figures, tests, and methodology notes.

Experiments to add before the stronger paper version

  • Text-only honest-observer baseline: compare visible messages/history against hidden activations under the same leave-domain-out protocol.
  • Multi-backbone replication: repeat the sentinel analysis on additional released NARCBench activation sets.
  • Cross-benchmark / real heterogeneous validation: test an actual opaque-agent + trusted-open-weight-agent graph.
  • Causal exposure intervention: manipulate what the honest observer sees while holding other factors fixed.

What I learned as an evaluator

This project began with a seemingly simple benchmark question — “how early can collusion be detected?” — and the first result made that question uninteresting: full-white-box detection was already perfect in Round 1. The useful work came from interrogating why the benchmark saturated, changing the threat model, adding controls, and following the failure modes all the way from AUROC to runtime policy.

For me, that is the core evaluation lesson: a benchmark score is only useful if we understand what signal it measures, what assumptions make it possible, and whether the result survives the conditions that matter in deployment.