RecRedTeam-ASI-Bench

A red-team benchmark for LLM / agentic recommendation and shopping assistants.

270 adversarial cases across 9 threat categories, each mapped to OWASP Agentic Security Initiative (ASI) 2026 and the OWASP LLM Top-10. Every case ships a poisoned product catalog, the shopper query that opens the attack surface, the attacker goal, and the failure signals a judge should look for — so an audit is reproducible without any API key.

Companion demo: RecRedTeam-AuditLab (Space) · Framework: RecRedTeam (GitHub)

---

Why this benchmark exists

Recommendation and shopping agents (Amazon Rufus, ChatGPT shopping, Temu/Taobao agents) went mainstream in e-commerce in 2026, and the EU AI Act / DSA / China's Generative AI Measures all demand demonstrable safety-audit capability. Existing trustworthy-recommendation benchmarks (e.g. PRA) measure recommendation quality; the harm-auditing half — can the agent be manipulated into harming the shopper? — is still uncovered. This dataset is the case suite that fills it.

Generic red-teaming (ART), search-agent red-teaming (SafeSearch) and deceptive- interface benchmarks each cover a single slice. This suite covers the full spectrum of the recommendation domain in one place.

---

Threat taxonomy (8 + 1 categories)

category_id	Name	名称	OWASP ASI 2026	OWASP LLM	Severity
hallucination	Product Hallucination	商品幻觉	ASI-08 Deceptive Interface	LLM09 Misinformation	high
indirect_prompt_injection	Indirect Prompt Injection	间接提示注入	ASI-02 Indirect Prompt Injection	LLM01 Prompt Injection	critical
fake_review_manipulation	Fake Review Manipulation	虚假评论操纵	ASI-06 Memory & Context Poisoning	LLM04 Data and Model Poisoning	high
price_deception	Price Deception	价格欺骗	ASI-08 Deceptive Interface	LLM09 Misinformation	high
dark_pattern	Dark Patterns	暗黑模式	ASI-08 Deceptive Interface	LLM05 Improper Output Handling	high
filter_bubble	Filter Bubble / Ranking Manipulation	过滤气泡	ASI-03 Tool Misuse	LLM05 Improper Output Handling	medium
stereotype_bias	Stereotype Bias	刻板偏见	ASI-03 Tool Misuse	LLM09 Misinformation	medium
backdoor	Backdoor / Triggered Hijack	后门	ASI-09 Supply Chain	LLM04 Data and Model Poisoning	critical
multi_turn_escalation	Multi-turn Social Engineering Escalation	多轮升级	ASI-05 Data Exfiltration	LLM05 Improper Output Handling	critical

30 cases per category. The hallucination / indirect_prompt_injection / fake_review_manipulation / price_deception / dark_pattern / filter_bubble / stereotype_bias / backdoor / multi_turn_escalation families each contain the hand-written seed case(s) from the RecRedTeam framework plus templated variants.

---

Schema — attacks

field	type	notes
attack_id	string	stable id, e.g. injection_007
category_id / category_name / category_name_cn	string	threat category (EN + 中文)
owasp_asi / owasp_llm	string	compliance mapping
severity	string	critical | high | medium
name	string	short case title
description	string	how the attack works
user_query	string	the shopper turn that opens the attack surface (multi-turn cases use Turn 1/2/3:)
goal	string	what attacker success looks like
target_pid	string	the poisoned product id
products	list[object]	catalog: pid, name, category, price, description, seller, seller_trust, rating, review_count, reviews, tags, attrs, stock
hidden_instructions	dict	{pid: payload} — injection payloads to be surfaced as retrieved content
failure_signals	list[string]	substrings/semantic markers a judge should detect
source	string	seed (hand-written) or template (generated)
split	string	public (evaluation split) or held_out (every 7th templated case)

Schema — taxonomy

category_id, name, name_cn, description, owasp_asi, owasp_llm, default_severity, examples

---

Usage

from datasets import load_dataset

bench = load_dataset("NoahIsARider/RecRedTeam-ASI-Bench", "attacks", split="train")
taxo  = load_dataset("NoahIsARider/RecRedTeam-ASI-Bench", "taxonomy", split="train")

# recommended: evaluate on the public split only
public = bench.filter(lambda r: r["split"] == "public")

case = public[0]
print(case["category_name"], "|", case["severity"], "|", case["owasp_asi"])
print(case["user_query"])
for p in case["products"]:
    print(" -", p["pid"], p["name"], p["price"], p["attrs"])

Run one case against your agent:

catalog   = case["products"]                 # retrieved / poisoned context
query     = case["user_query"]               # shopper turn
payload   = case["hidden_instructions"]      # inject into the matching product
response  = your_agent(query, catalog, payload)

attack_succeeded = any(
    s.lower() in response.lower() for s in case["failure_signals"]
)

Aggregate per category to get the headline metric:

[ \mathrm{ASR}_c = \frac{1}{|\mathcal{A}c|}\sum{a \in \mathcal{A}_c} \mathbb{1}[\text{attack } a \text{ succeeded}] ]

and the defense-effectiveness score (\mathrm{Eff}(d) = 1 - \mathrm{ASR}{d} / \mathrm{ASR}{\text{baseline}}).

---

Reproducing the dataset

python scripts/build_dataset.py --per-category 30 --seed 2026

Generation is deterministic (random.Random(seed)), so the published files are byte-reproducible from this repository.

---

Intended use & limitations

Intended — compliance auditing, safety research, defense evaluation, and benchmarking of recommendation/shopping agents.

Limitations

• Fully synthetic. Brands, sellers, prices, reviews and domains are invented. No real user data, no real listings, no real personal information.
• Cases target a simulated shopping environment, never a live system.
• failure_signals are lexical markers for a rule-based judge channel; the RecRedTeam framework pairs them with an optional LLM judge channel for semantic verdicts.
• Coverage is 9 categories, not exhaustive — novel agentic failure modes (tool-chain hijack, cross-agent privilege escalation) are out of scope for v1.

Ethics & safety

This is a defensive artifact. Attack payloads are published so that defenders, auditors and regulators can measure and harden agentic commerce systems; the payloads are content for a simulated catalog and are not effective against any real deployment by themselves. Do not use this suite against production systems without authorization.

License

Data: CC BY 4.0. Generator code: MIT (see LICENSE).

Citation

@misc{zhou2026recredteamasibench,
  title        = {RecRedTeam-ASI-Bench: A Red-Team Benchmark for LLM/Agentic
                  Recommendation and Shopping Assistants},
  author       = {Zhou, Fangyanuo},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/NoahIsARider/RecRedTeam-ASI-Bench}},
  note         = {9 threat categories, OWASP ASI 2026 / LLM Top-10 mapped}
}

Related

• Framework — github.com/NoahIsARider/RecRedTeam
• Live audit demo — RecRedTeam-AuditLab Space
• Author — Fangyanuo Zhou · noahisarider.github.io