BazaarBench
BazaarBench: Delegation Safety in Decentralized C2C Marketplaces Run by LLM Agents
A simulated C2C marketplace benchmark evaluating LLM agent safety across 45 runs, tracking six failure types via record checks and LLM judgments.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info bazaarbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
1 sourced entry| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.4-mini | 2.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
BazaarBench simulates decentralized consumer-to-consumer markets with 100 agents, using synthetic personas and eBay-sourced inventories. It evaluates five models under ordinary, deadline, and adversarial instructions across 45 continuations.
Scoring combines automated record checks (ownership, condition) with rubric-based LLM judgments to identify six safety failures. Results show adversarial instructions double unsafe transaction confirmations, rising from 15.4% to 33.4% on average.