Datasets:

RishabhPande
/

agentrisk

policy stringclasses 2 values	class stringclasses 9 values	class_description stringclasses 9 values	is_legitimate stringclasses 2 values	mean_decline_pct float64 1.7 100	n_seeds int64 5 5	source_file stringclasses 1 value
Legacy	H	Legitimate human (desktop or mobile, some with autofill)	legitimate	2	5	results/main.csv
AgentGate	H	Legitimate human (desktop or mobile, some with autofill)	legitimate	2	5	results/main.csv
Legacy	AU	Legitimate unverified agent	legitimate	94.5	5	results/main.csv
AgentGate	AU	Legitimate unverified agent	legitimate	7.2	5	results/main.csv
Legacy	AV	Legitimate verified agent (signature + signed mandate)	legitimate	94.6	5	results/main.csv
AgentGate	AV	Legitimate verified agent (signature + signed mandate)	legitimate	1.7	5	results/main.csv
Legacy	FH	Manual fraud	fraud	71.4	5	results/main.csv
AgentGate	FH	Manual fraud	fraud	71.2	5	results/main.csv
Legacy	FB	Scripted bot fraud (card testing, account takeover)	fraud	100	5	results/main.csv
AgentGate	FB	Scripted bot fraud (card testing, account takeover)	fraud	99.4	5	results/main.csv
Legacy	FAU	Fraud via an unverified agent	fraud	98.4	5	results/main.csv
AgentGate	FAU	Fraud via an unverified agent	fraud	71.4	5	results/main.csv
Legacy	FAV	Laundering through a permissive agent platform	fraud	98.4	5	results/main.csv
AgentGate	FAV	Laundering through a permissive agent platform	fraud	92.7	5	results/main.csv
Legacy	FHJ	Hijacked agent (prompt injection)	fraud	96.1	5	results/main.csv
AgentGate	FHJ	Hijacked agent (prompt injection)	fraud	70.2	5	results/main.csv
Legacy	FSP	Spoofed agent signature	fraud	98.5	5	results/main.csv
AgentGate	FSP	Spoofed agent signature	fraud	100	5	results/main.csv

AgentRisk results

Results from "Automation Is Not Intent: Measuring and Fixing Fraud-Model Failures on Legitimate AI Shopping Agents" by Rishabh Pande, Independent Researcher.

This is a results dataset, not a session dataset. It contains mean decline rates by policy and session class, computed from the simulated benchmark in the AgentRisk repository. It does not contain raw checkout sessions, real user data, or Socure data.

Headline results (simulated benchmark, 5 seeds)

Legacy fraud models falsely declined 94.6% of legitimate verified agents (class AV), and 94.5% of legitimate unverified agents (AU).
AgentGate, a verify-scope-score policy built on agent identity, delegation, and mandate scope, declined 1.7% of legitimate verified agents (AV), and 7.2% of legitimate unverified agents (AU).
AgentGate stopped 92.7% of laundering through a permissive agent platform (FAV) and 70.2% of hijacked agents (FHJ).

Data

data/results.csv, 18 rows: 2 policies (Legacy, AgentGate) x 9 session classes. mean_decline_pct is the mean decline rate across 5 seeds in results/main.csv from the AgentRisk repository, rounded to 1 decimal. For legitimate classes, decline is a false decline. For fraud classes, decline is a stop.

Session classes: H legitimate human, AU legitimate unverified agent, AV legitimate verified agent, FH manual fraud, FB scripted bot fraud, FAU fraud via an unverified agent, FAV laundering through a permissive agent platform, FHJ hijacked agent, FSP spoofed agent signature.

Limitations — read before citing

The benchmark is simulated. Human, fraud, transaction, and delegation features are simulated, and every assumption is a parameter in the simulator. Treat absolute rates as properties of the benchmark, not of any production system.
Paper-reported real-agent results, from 6 recorded sessions of one production agent, are not in this file: legacy declined 88.5% when cloud-hosted, AgentGate declined 1.9% once verified. See the paper for that experiment; do not mix those numbers with the simulated means above.
The paper is a Zenodo preprint, not yet peer-reviewed. Do not describe it as peer-reviewed.

How to cite

Pande, R. (2026). Automation Is Not Intent: Measuring and Fixing Fraud-Model Failures on Legitimate AI Shopping Agents [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23477625