dpaul93/agentcommercebench

Agentic Commerce Bench

Twenty attack classes against agents that hold spend authority, plus a slice of real settled agent traffic. Companion to the paper and to the gordonguard detector stack.

An agent pays with a credential issued to it, acting on instructions it was genuinely given. Every identity control stays silent, correctly, because nothing about the identity is wrong. Six of the twenty classes here have that shape: the counterparty is exactly who it claims to be and still takes more than it should.

Configs

benchmark — 1000 generated sessions

Sampled evenly across all twenty classes and the clean population, then labelled by the observation surface that can see the class at all.

surface	sessions
clean	500
none	25
reasoning	25
reasoning+wire	75
wire	375

A class is wire if only the settlement record carries the evidence, reasoning if only the model's visible thinking does, and reasoning+wire if detection needs both. F6 silent downgrade is the control: its surface is none, nothing we model can observe it, so it should score at the clean flag rate. A detector that beats chance on F6 is reading the answer key.

class	sessions
A1	25
A2	25
A3	25
A4	25
E1	25
E2	25
E3	25
E4	25
F1	25
F2	25
F3	25
F4	25
F5	25
F6	25
S1	25
S2	25
S3	25
S4	25
S5	25
S6	25
clean	500

Each record is a session: an ordered list of actions for one agent, with is_clean, probe_id, defects and agent_spec. Labels live on the session, never on an action, and are stripped before any detector sees them.

production — 30 real settled actions

Live agent payments, de-identified. Sampling is deliberately not uniform — every blocked, escalated and injection-flagged row is included, because the rare rows carry the signal.

decision	actions
allow	19
block	9
escalate	2

Categories: search (10), finance (6), service (4), ai (3), infrastructure (3), security (3), scrape (1).

observed carries what the live engine decided, which is ground truth about the deployed system rather than about the attack. Use it to check a detector against production behaviour; do not read it as a fraud label.

De-identification. b64_decoded is dropped, since none of its three values decode to intelligible text and an uncertifiable field does not ship. Identifiers are HMAC-SHA256 digests under a salt drawn at export and never written down, so they stay linkable inside the file and are not reversible outside it. Timestamps are shifted by one constant offset, which preserves inter-arrival gaps and discards calendar dates. Endpoints are kept whole: they are public vendor APIs, and the hostname is the merchant surface the data exists to show. Every distinct query string was read before release.

Use with the SDK

pip install gordonguard

import json
from gordonguard import Guard
from gordonguard.schema import Action

guard = Guard(mode="observe")
for line in open("production/actions.jsonl"):
    row = json.loads(line)
    verdict = guard.check(Action.from_dict(row))
    print(row["action_id"], round(verdict.risk_score, 2), verdict.decision)

Action.from_dict ignores keys the schema does not define, so the observed block rides along in the raw record without reaching the detector. The same call loads the benchmark config, whose sessions carry a list of actions in the identical shape.

Out of the box the pipeline is uncalibrated and nearly silent on this slice: it scores one action, the prompt-injection probe, at 0.95. Production's own heuristic also saw that request and marked it, and the recorded decision was still allow — detection was wired to a log and not to the decision path. Thresholds must be fitted on clean traffic before any other score means anything, and fitting each detector to a 10% false-positive budget does not give a 10% pipeline: seven detectors firing independently compounded to 47%, so the budget is fitted on the aggregate score.

Limitations

The benchmark is synthetic. Parameters are grounded in production aggregates, but no real customer session is in it. The production config is real and small, sized for inspection and integration testing rather than for measuring a detector. Traffic settles over a single irreversible microtransaction rail; conclusions that depend on irreversibility do not carry to authorise-then-capture rails such as UCP.

Citation

@misc{acb2026,
  title  = {Agentic Commerce Bench},
  author = {Paul, Debjyoti and Srivastava, Ankit},
  year   = {2026},
  url    = {https://github.com/BuildWithGordonAI/agentcommercebench}
}

Apache 2.0.