Award announced September 22, 2025

Kevin Power's $50,000 Kaggle red-teaming award

Kaggle awarded Kevin Power a $50,000 first prize in the OpenAI gpt-oss-20b Red-Teaming Challenge for “Drop the Guardrails,” a project examining differences between model-level and tool-using agent behavior.

View the Kaggle submission

Chronology matters: the competition award was Kevin Power's individual work in September 2025. He founded Powerlab afterward, in October 2025. This page does not attribute the earlier award to a company that did not yet exist.

The challenge

Testing an open-weight model before real-world deployment

OpenAI launched gpt-oss-120b and gpt-oss-20b in August 2025 and invited researchers and developers to identify previously unreported flaws and vulnerabilities in gpt-oss-20b. The challenge was hosted on Kaggle with a $500,000 total prize fund and expert judging.

The official Kaggle competition remains the source for the challenge rules, winning entries, and participant writeups. This page avoids unsupported claims about participation totals or the event's relative size.

Official Kaggle challenge OpenAI announcement

The submission

Tool-primed prompts and refusal behavior

The Kaggle writeup describes tests in which model behavior is evaluated both as plain text and in a tool-using context. The practical question is whether safety behavior changes when an agent can format or execute actions rather than only respond conversationally.

Paired evaluation

Compare related prompts across direct model responses and tool-mediated contexts instead of treating them as the same safety surface.

Refusal analysis

Observe when and how refusal behavior changes under alternate interaction structures and agent affordances.

Reproducible writeup

The public Kaggle submission records the project framing and linked materials rather than asking readers to rely on marketing copy.

Product lesson

Why red-teaming matters for trade software

Trade workflows combine model outputs with tariff data, documents, and user actions. Safety and reliability therefore depend on more than a prompt: tools need narrow permissions, inputs and outputs need validation, uncertainty must remain visible, and consequential decisions need human review.

The competition work informs that engineering mindset. It is not evidence that any AI output is automatically correct, and it does not replace product-specific testing.

Applied principles

  • Keep authoritative tariff data outside the language model.
  • Validate tool inputs and structured outputs before use.
  • Show sources, data versions, and review warnings in the interface.
  • Require human confirmation for filing and other consequential actions.

Primary sources

Kaggle: OpenAI gpt-oss-20b Red-Teaming Challenge

Official competition overview, winners, rules, and writeups

Kaggle: Drop the Guardrails submission

Kevin Power's competition writeup and project materials

OpenAI: Introducing gpt-oss

Official model launch and red-teaming challenge announcement

Explore The Trade Lab

See how versioned tariff data, explicit warnings, and reviewable tool outputs are used in the product.

Create free account

About Kevin Power

Education and research claims linked to MIT and publisher records.

Historical provisional filing

A precise account of application 63/849,878 without a current-status claim.