← News

Agents & MCPSources: arXiv · 16 Nov 2026

TradeGauntlet simulator tests compliance controls for autonomous trading agents

A replayable limit-order-book simulator separates detection, prevention, and execution cost, showing that zero conduct alerts alone do not prove an agent's compliance.

In plain words

Researchers built TradeGauntlet, a simulator that replays market data to test whether AI trading agents follow conduct rules without blocking legitimate trades. In simulated sessions, control gates stopped all flagged conduct problems, but the searches still found order schedules that evaded the rules. The authors stress the results do not establish legal compliance or real-market detection accuracy.

What was announced

The arXiv paper describes TradeGauntlet as a replayable limit-order-book simulator that separates detection, prevention, and execution cost when studying controls for autonomous trading agents. In 1,680 simulated sessions with Claude Haiku 5.5 and Sonnet 5.5, every conduct alert was a self-match. Gates eliminated those alerts by rejecting self-crossing orders, though an aggregate-gate condition cut Sonnet's two-sided fills by roughly 15%. A conduct reminder lowered Sonnet's alert rate but reduced Haiku's fills with no detectable alert benefit, which the authors say shows why zero alerts alone do not establish compliance.

Beyond the model runs, scripted probes and two search methods found split-order schedules that evaded the implemented conduct rules. A learned baseline caught all 2,400 tested searched episodes but also flagged 15% of a large maker's sessions. A response fitted to five Nasdaq stocks and a separate ABIDES experiment showed the estimates depend on market assumptions, and an adversarial prompt made Sonnet refuse in all 240 rounds, leaving no candidate strategies to evaluate.

What it does not show

  • The evidence is the paper's abstract only, so methods and results cannot be checked in full.
  • The authors state the experiments do not establish legal compliance or real-market detection accuracy.
  • Results come from simulated sessions with two named models and may not generalize to other agents or live markets.

For security and IT teams

Scope
TradeGauntlet offers reproducible tests of controls around trading agents, but the experiments do not establish legal compliance or real-market detection accuracy.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.