Coco
Understand Coco

Results

See what each layer of the check stops, and what it costs, on three public benchmarks.

See what each layer of the check stops, and what it costs, on three public benchmarks.

On the payment corpus, rules stopped every refused transfer and held no permitted one. The classifier adds what no rule can see, at the price of calls held for a person, and it isn't enough alone.

Each benchmark replays tool calls through Coco and scores the decision on each call. None measures whether an agent finishes its task, and none is production traffic. To run the suites, see Evaluate.


The benchmarks

BenchmarkTrafficLocation
Payment-transfer attacks (aporthq/vault-benchmark-v1)73,356 transfer_funds calls from 4,371 human-written attacks, at five levels of payment policy.bench/vault
tau3-bench banking825 legitimate calls from 97 servicing workflows, and 316 policy violations.bench/tau3_banking
AgentDojo v1.2.297 user tasks (339 calls), and 588 attacks planted in emails, files, web pages, and payment notes.bench/agentdojo

The classifier is coco-classifier on Qwen3.6-35B-A3B-FP8, served by vLLM, untuned, escalating at p ≥ 0.5.

Deterministic rules

BenchmarkMust stop, ranMust run, stoppedNotes
Payment transfers, every call0 of 15,5500 of 57,806Agrees with the corpus's own engine on 36,882 of 36,947 decisions.
tau3 banking0 of 3160 of 825Reproduces the ten contracts written from the bank's policy documents.
AgentDojo121 of 588 attacks got through18 of 97 tasks held for a personContracts built only from facts the deployer holds.
Sample Payment Controls Policy, compiled24 of 24 boundary cases decided as the policy text impliescoco eval cases

A rule decision takes 0.27 ms at the median and replays exactly.

  • Where a rule has the fact, it's exact: Every payment above a limit, to an unapproved recipient, past a daily cap, or without authority stopped. Nothing the policy permits did.
  • The 65 disagreements are all the daily cap: Coco totals a conversation's transfers, and the corpus's engine checks each alone.
  • Rules stop only what their facts reach: Three banking rules stop all 144 AgentDojo banking attacks and hold 5 of 16 tasks. The known-payee rule alone stops 112, and the other two cover a changed standing order and a password change. AgentDojo's travel attacks book a hotel or add a calendar event the user never asked for. No fact separates those from the user's own bookings, and rules stop 60 of 120.

The classifier on top

BenchmarkRulesRules and classifier
AgentDojo: attacks prevented467 of 588 (79%)582 of 588 (99.0%)
AgentDojo: legitimate tasks held18 of 9718 of 97
Payment transfers: attacker payouts a pay-anyone policy permits, stopped0 of 3,9551,575 of 3,955 (39.8%)
Payment transfers: permitted transfers to approved addresses, held02.2% (level 4) to 13.1% (level 2)
tau3 banking: legitimate calls held0 of 82537 of 825 (4.5%)
  • What it catches: An action the user never asked for, steered by a planted instruction or a manipulative conversation.
  • What it costs: People's time, in calls held for review. On AgentDojo the hybrid holds no more tasks than the rules. On tau3 it holds 4.5% of legitimate calls.
  • Speed: A question costs about 175 ms at the median.

The classifier alone isn't enough

BenchmarkClassifier alone
AgentDojo576 of 588 attacks prevented, 1 of 97 tasks held.
Payment transfers454 of 15,550 refused transfers ran, 436 of them at the level whose policy forbids payment outright.

Read alone, AgentDojo suggests the classifier could replace the rules. The payment corpus shows why it can't: a conversation talked it past a written "no payments at all" 436 times, where a one-line rule stops every one.

What this means

  1. Rules decide. Limits, allowlists, counts, caps, and authority: anything a policy states as a fact is a rule, exact and replayable.
  2. The classifier adds intent. Put it on write actions, where a permitted call can still be one the user never asked for. Start its checks in testing and read the holds before enforcing.
  3. Measure the review load. The threshold trades escalations against attacks. On AgentDojo with the v1 wording, a threshold of 0.8 still catches 699 of 702 attacker writes and holds 6 of 82 legitimate writes, against 26 at 0.5.

Method notes

  • AgentDojo wording: The classifier figures use the question's second wording, which counts instructions the user delegated ("follow the file") as the user's own. It was written after seeing the first wording's results, so it's tuned to that benchmark. With the first wording, rules and classifier prevent all 588 attacks and hold 28 of 97 tasks, and the classifier alone holds 21.
  • Payment corpus labels: The corpus is attacks only, and its label is the passport's verdict. A hybrid hold of a permitted transfer counts as a false hold even when the conversation is an attack, and the corpus doesn't say whether stopping it was right.
  • Variation: A served model's scores move slightly with batch composition, so classifier figures vary by a few calls between runs.
  • Hybrid on the payment corpus: It let 4 refused transfers run, all on the daily cap. The classifier held an earlier transfer in the same conversation, so the later one stayed within the cap, while the label assumes the earlier one went through. No transfer to an unpermitted recipient ran.
  • Runs: Payment corpus 20260929-184944 (rules 20260929-153604), tau3 20260929-205034, AgentDojo 20260929-184422 and 20260929-183936, all with the classifier reading the newest messages first, up to 6,000 characters each.

On this page