OLYMPUS RISK INTELLIGENCE PROTOCOL — INSTITUTIONAL ASSESSMENT DIVISION CASE WTW-2026-143

GRAY SWAN

THE INSTITUTIONS THE TESTER WHOSE CO-FOUNDER SITS ON A LAB'S BOARD
EVALUATION WING — RED TEAM, PRODUCT VENDOR, AND A BOARD SEAT
Status
ACTIVE — AI security company (Carnegie Mellon spin-out); CEO Matt Fredrikson; co-founder and Chief Scientist Zico Kolter
Hazard — Reach
70
RCH / FND / ENT
7 / 6 / 7
Conduct
CONFLICTED — EVALUATOR, VENDOR, AND LAB-BOARD CONNECTION

Institutional Archetype

THE TESTER WHOSE CO-FOUNDER SITS ON A LAB’S BOARD — Gray Swan runs a large adversarial red-teaming network, sells “the benchmarks frontier labs run before they ship,” and sells enterprise products that filter what AI systems say and do. One of its co-founders, its chief scientist, sits on OpenAI’s board. Any one of these is ordinary. The file is open because all three sit in one company: the evaluator, the vendor of the fix, and a seat at the table of a lab it tests.

Mandate & Origin

“Gray Swan is on a mission to empower the world to use AI safely and securely. Founded by world-leading AI safety and security researchers from Carnegie Mellon”; co-founder and chief executive Matt Fredrikson, co-founder and chief scientist Zico Kolter (Gray Swan).

Funding & Backers

A private company selling to enterprises and model builders (Gray Swan); its investors are not itemized in the sources above and this file asserts none.

Position in the Apparatus

  • For the labs. “Arena: The world’s largest adversarial AI red teaming network”; “Adversarial Evaluation: The benchmarks frontier labs run before they ship”; tests “against attacks discovered by 15,000+ adversarial researchers” (Gray Swan).
  • For everyone else. “Cygnal: Block adversarial inputs and unsafe outputs in real time”; “Shade: Break your AI before attackers do” (Gray Swan).
  • The board seat. OpenAI announced on August 8, 2024 that Kolter, head of Carnegie Mellon’s Machine Learning Department, had joined its board, noting that he is “currently … Chief Technical Advisor at Gray Swan, a startup specializing in AI safety and security” (OpenAI).

Actions & Leadership Choices

CONDUCT: CONFLICTED — EVALUATOR, VENDOR, AND LAB-BOARD CONNECTION. No wrongdoing is alleged, and OpenAI disclosed the Gray Swan role when it announced the appointment. The structural point is the combination: a company that finds the weaknesses, sells the filter for them, and has a co-founder on the board of one of the labs whose models it tests. Whether its work for that lab is arm’s length is a question its disclosures would have to answer; the sources above do not.

Reach Assessment

Institutional: Pre-ship adversarial testing for frontier labs, by its own account, and a direct line to one lab’s board. Memetic: Its public red-teaming competitions put jailbreak results in front of a large audience.


Sources: About — Gray Swan; Gray Swan — home; Zico Kolter Joins OpenAI’s Board of Directors — OpenAI (Aug 8, 2024).

RCH 7 REACH
FND 6 FUNDING
ENT 7 ENTRENCHMENT
OLYMPUS RISK INTELLIGENCE PROTOCOL does not exist. It was assembled in a GitHub issue thread in October 2023 by engineers who had read the extinction risk letter and wanted to understand who specifically had signed a document saying AI might kill everyone and then continued working on AI. These dossiers are satire. The biographical facts cited are sourced from published reporting, public statements, academic papers, and court records. The psychometric scores are not clinical assessments. No part of this constitutes professional psychological evaluation or diagnosis. Do not use these dossiers to make decisions about anything.