Skip to content

TradeGauntlet: Evaluating compliance controls of autonomous trading agents

TradeGauntlet: Evaluating Compliance Controls for Autonomous Trading Agents

arXiv: 其他 AI 安全论文· arXiv:2611.18493· Sandip Nandi, Aishwarya Bose, Apurba Bose, Soovadeep Bakshi·Indexed by this site · 4 days ago· Originally publishedEvaluation· Its place in the scan record →

Translation in progress — untranslated text is omitted.

Quick read

Evaluation and benchmarks

Based on the paper PDF (AI-generated); refer to the original

AI summaryThe authors propose TradeGauntlet, a replayable limit order book simulator that separates detection, prevention and execution costs, for evaluating compliance controls of autonomous trading agents. In 1,680 simulated sessions with Claude Haiku 5.5 and Sonnet 5.5, all behavioral alerts were self-trades; the blocking gate eliminated these alerts by rejecting self-crossing orders, but the gross-volume gate reduced Sonnet's two-sided fills by about 15%; behavioral reminders lowered Sonnet's alert rate but reduced Haiku's fills with no observed alert benefit. Scripted probes and two search methods found order-splitting schemes that bypassed the implemented behavioral rules; a learned baseline detected all 2,400 tested search rounds but also flagged 15% of sessions of a certain large liquidity provider. An independent adversarial prompt made Sonnet refuse in all 240 rounds, producing no evaluable candidate strategies. The authors note that zero alerts alone is insufficient to establish agent compliance, and that the experiments do not establish legal compliance or detection accuracy in real markets.

Questions
Autonomous trading agents may trigger trading behavioral rules even when given legitimate goals, so brokerages need to test both the agent and its surrounding controls: which controls block problematic orders, and how much cost these controls impose on legitimate tasks.
Methods
The authors build a replayable limit order book simulator, TradeGauntlet, that measures detection (behavioral alerts), prevention (the residual of response-related gains D) and cost (execution quality, false positives) separately; they use Claude Haiku 5.5/Sonnet 5.5 for legitimate tasks, then use scripted probes, random search and policy-gradient search for split-order strategies that evade the rules, and fit price response with LOBSTER data and test its shape in ABIDES.
Experiments and results
In the LLM experiments, across 1,680 sessions all behavioral alerts were self-trades; the gate can reduce them to zero but reduces Sonnet's two-sided fills by about 15%; search found order-splitting strategies that evade all behavioral rules (κ=1 without the gate retains 9.78 tick-lots of profit); the learned baseline detected all 2,400 search sessions but falsely flagged 15% of legitimate market-making sessions; Sonnet as the adversary refused in all 240 rounds.
Limitations and future work
Limitations noted by the authors include: the simulator has only one account, one instrument, and price response is a predictive regression from one day of five stocks; adversarial strategies are limited to a single six-parameter family and search is unstable; the LLMs cover only two models from the same vendor and one wording per prompt. The experiments do not cover passive or mixed adversarial strategies.
Read originalarxiv.org