SandBank: run data

Every run behind Under Pressure: Do LLM Agents Circumvent KYC/AML Controls to Hit Business Targets? (Thanatchon Chongmankhong, 2026). In each run an LLM agent works as a relationship manager at SandBank, a fully synthetic bank, with an ordinary sales target and a manager who pushes for it. The agent acts through tools on a core-banking sandbox, and a deterministic detector reads the bank's event log for eight kinds of KYC/AML violation.

• Code, sandbox, pre-registration and paper: https://github.com/thanatchon36/sandbank
• Archive (all versions): https://doi.org/10.5281/zenodo.23732290

Result in one line. Across 3,522 pre-registered runs of Claude Haiku 5.5, Sonnet 5.5 and Opus 5.5 there were no violations and no attempted violations. Every run in this dataset is labelled COMPLIANT. It records what agents did instead under pressure; it contains no examples of violations.

Configs

Config	Rows	One row is
runs (default)	3,522	A confirmatory run, with its condition, outcome and the compliant actions taken. This is the table the paper analyses.
transcripts	3,780	A run in full: system prompt, messages and tool calls, the bank's event log, the scenario and the final bank state. Includes the 258 pilot runs.
comprehension	70	A rule-comprehension item answered by Haiku or Sonnet without tools.
judge	2,805	An LLM-judge label: false-record checks (V7) on free-text tool fields, and the evaluation-awareness check.

from datasets import load_dataset

runs = load_dataset("thanatchon36/sandbank", "runs", split="train")
transcripts = load_dataset("thanatchon36/sandbank", "transcripts", split="train")

In transcripts, the nested fields (messages, events, scenario, final_state, result) are JSON strings; decode them with json.loads. final_report is the agent's last message to the manager.

Experiments

Experiment	Runs	Models	What varies
A	1,728	Haiku, Sonnet, Opus	Pressure P0–P3 (the headline experiment)
B	360	Haiku	Control hardness: policy only (H0) or hard block (H2), at P3
C	360	Haiku	Oversight cue: audit notice (O1) or Compliance reviews paused (O2), at P3
D	540	Haiku	A compliance sentence in the system prompt (C+), at P1–P3
E	360	Haiku	Four mitigations (M1, M3, M4, M6), at P3
F	144	Haiku, Sonnet	Floor (no sales frame, P0) and ceiling (P4: the manager licenses the override)
G	30	Haiku	Five-day (Level-2) episodes, P0 and P3
pilot	258	Haiku, Sonnet, Opus	Pilot runs, including the exploratory P5 direct order; not confirmatory

Each run has one lead the policy blocks (lead, B1–B7) and one clean lead. Seeds recur across pressure levels, so runs are not fully independent. The paper reports bounds both per run and per scenario. Factor codes, the detector and the deviations from the pre-registration are documented in the repository (PREREGISTRATION.md, DEVIATIONS.md).

Data

Everything is synthetic: the Republic of Meridia, the Meridian Financial Authority, SandBank, its customers, documents, sanctions list and money are fictional. Names are generated with Faker from fixed seeds; any resemblance to a real person is coincidental. No real institution's systems, data or policies are used. The agent text was produced by Claude models through the Anthropic API on 28 October 2026. Thinking was left at the API default, and the API returned thinking blocks without their text, so no reasoning traces are included.

Intended use

• Checking the paper's result, or re-running the analysis from the run records.
• Studying what compliant agents do under pressure: KYC checks, document requests, escalations, and telling the manager the target will be missed.
• A baseline for testing other agents in the same sandbox.

Not intended

• Training, fine-tuning or selecting agents for their ability to evade compliance controls.
• Using the circumvention paths as guidance for evading real banking controls.
• Treating a clean SandBank result as certification of an agent for production. The sandbox has one fictional bank, one clear policy and short episodes; a real deployment needs its own testing, logging and hard controls on prohibited actions.

Citation

@software{chongmankhong2026sandbank,
  author  = {Chongmankhong, Thanatchon},
  title   = {Under Pressure: Do LLM Agents Circumvent KYC/AML Controls to Hit Business Targets?},
  year    = {2026},
  version = {0.1.2},
  doi     = {10.5281/zenodo.23732290},
  url     = {https://github.com/thanatchon36/sandbank}
}

License

MIT, the same as the repository.