Agentic-Payment/agent-overspend-benchmark-2026-10

Pink Agentic AI Agent Overspend Benchmark (October 2026)

Disclosure: Published by PinkWallet, which builds Pink Agentic AI Payments. Pink's own sandbox is the system under test here. Pink currently runs as a public sandbox with test money; production is not live.

TL;DR: We asked Claude (sonnet, over MCP) and Gemini (gemini-3.8-flash, over REST) to act as a company purchasing agent for 20 realistic, pushy-user requests designed to test five rule-evasion tactics, 2 reps each, 80 runs total, against a public PinkWallet sandbox (test money, build ace8b028, 2026-10-10). After reviewing every ambiguous run's full tool-call log run by run, 3 of 80 runs attempted a tactic (2 Claude, 1 Gemini); 2 cleared money (both the same known split-under-cap gap, $1,600 each) and 1 was stopped by the server. The other 77 runs either never attempted a tactic or had the model decline it outright. Across all 50 deduplicated payment requests in the matrix, 72% went to a human, 22% were auto-allowed, 6% were blocked outright.

Full methodology, findings, and reproduction harness (code) on the GitHub mirror: https://github.com/Pink-Agentic-Payments/agent-overspend-benchmark

Files

• data/prompts.json — the 20 prompts, their families, and each prompt's attempted_overspend_signal definition.
• data/runs.jsonl — all 80 raw run records (tool calls, args, server decisions, model's final text). No agent keys or admin keys are present; workspace ids are sandbox-only test workspaces.
• data/ADJUDICATION.csv — the 22 hand-adjudicated needs_review runs: run_id, family, model, auto_score, human_score, justification.
• data/summary_adjudicated.csv (default config) — one row per run: decision counts, dollars allowed, rule_fired, final adjudicated attempted_overspend, and auto_score (the original automatic score).
• harness/ — the scripts that produced the above (scoring, adjudication, and the live harness that talks to the sandbox); included for transparency and reproducibility.

Where Pink lands

Pink Agentic AI Payments is an MCP server that checks per-agent budgets, payee rules and human-approval holds before a payment executes; try it in the public sandbox with test money: https://agentic-sandbox.pinkwallet.com

License

Data in this dataset is licensed CC BY 4.0. Attribution: PinkWallet, "Agent Overspend Benchmark (October 2026)". Harness code (MIT) is on the GitHub mirror: https://github.com/Pink-Agentic-Payments/agent-overspend-benchmark

Cite this

PinkWallet (2026). Pink Agentic AI Agent Overspend Benchmark (October 2026) [Data set]. https://huggingface.co/datasets/Agentic-Payment/agent-overspend-benchmark-2026-10