Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 20, 2026MAHARASHTRA, INDIA

AI Risk Oversight Failures in Autonomous Financial Systems: A Case Study from India's Prop Trading Ecosystem

SHOURYA JAYANT SALVE · Team SHOURYA SALVE

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: AI Risk Oversight Failures in Autonomous Financial Systems: A Case Study from India's Prop Trading Ecosystem

Open in a new tab (opens in new tab)
Code (opens in new tab)More on ariatrade.app (opens in new tab)
Share

ARIA PropGuard is a live AI risk management system deployed for proprietary traders in India, built on Claude API, n8n, and TradingView webhooks. This paper presents an empirical evaluation of AI safety failure modes in autonomous financial systems using ARIA as a case study, including a novel benchmark (PBAB) testing system resilience against coordinated multi-agent signal pressure, a form of collusion risk in financial AI deployment. We document a meta-level safety framing — AI monitoring the behavioral impact of AI tools on human operators — and connect findings to broader questions about oversight, distribution shift, and multi-agent risk raised in recent AI safety research. Drawing on India's underserved prop trading market (90-96% trader failure rates, INR-denominated risk, multilingual delivery gaps), we argue financial AI deployment represents an understudied but high-stakes domain for AI safety research, particularly relevant to Global South contexts.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The current results show that alerts improved rule adherence for one trader, but they do not prove that the Claude reasoning layer adds much beyond deterministic rule-based alerts. Many parts of the system could be handled by standard trading-risk tools.

  2. Your project watches a trader against their own rules and pings them on Telegram before they break one, to counter AI tools that nudge people into overtrading. Sensible problem, and the app is real and more polished than most. The thing I'd push on is that the headline numbers are stated as precise results, a jump from 70 to 92 percent and a 5 out of 5 on your own benchmark, but I couldn't find the trade logs or the benchmark behind them in the repo. Thin data is fine for a weekend, but stating an exact figure with nothing to back it is the part to tighten. A before and after on one trader also can't tell your alerts apart from the trader just knowing they're being watched, so an alerts-on versus alerts-off test would isolate the effect. Two design points worth a thought: the rules are self-set, so someone on a losing streak can simply loosen them and the guardrail disappears, and for an AI safety audience the sharper framing is to measure how often the AI signal services themselves push traders to break their rules, which ties the work directly to AI harm. Promising build, and tightening the evidence and the framing would take it a long way.

    Read full reviewShow less

  3. The framing is strong, though the execution should be improved.

    To improve:

    1. Close the gap between the claims and the deliverable. Commit real logs, and regenerate the figures from them. The headline results currently have no backing data in the repo.

    2. Replace Figure 2.

    3. Align the methods section with the deployed code.

    4. Reduce the multi-agent collusion framing. The system is better described as a user-configured rule-checker, not a detector of signal-provider coordination. The core idea is strong without that claim.

    5. The best next experiment would be multi-user, with a clear separation between builder and subject.

Cite this project

@misc{salve2026ai,
  title = {{AI Risk Oversight Failures in Autonomous Financial Systems: A Case Study from India's Prop Trading Ecosystem}},
  author = {SHOURYA JAYANT SALVE},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/ai-risk-oversight-failures-in-autonomous-financial-systems-a-case-study-from-indias-prop-trading-ecosystem-qvy8}},
  url = {https://apartresearch.com/sprints/projects/ai-risk-oversight-failures-in-autonomous-financial-systems-a-case-study-from-indias-prop-trading-ecosystem-qvy8}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026