MATT FREDRIKSON

Behavioral Archetype

THE PROFESSOR WHO BROKE THE GUARDRAILS, THEN SOLD NEW ONES — Subject co-wrote the 2023 paper showing that one automatically generated string could jailbreak ChatGPT, Bard and Claude at once. He now runs the company that red-teams frontier models before release and sells filters against the attacks it finds. His co-founder chairs OpenAI’s Safety and Security Committee. This profile scores the position, not the man.

Essence Indicators

Co-founder and chief executive, Gray Swan, with co-founder and chief scientist Zico Kolter (Gray Swan).
Associate Professor, Carnegie Mellon University; research interests “Formal Methods” and “Security and Privacy” (Carnegie Mellon).
The attack. Co-author, with Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr and Kolter, of Universal and Transferable Adversarial Attacks on Aligned Language Models (July 2023) (arXiv 2307.15043).

Forensic Archetype Comparison

Pattern	Match Level	Evidence
The Attacker-Vendor	HIGH	Wrote the attack; sells the evaluation and the filter.
The Academic Founder	HIGH	Professor and CEO at once.
The Board Proximity	MODERATE	His co-founder chairs OpenAI’s Safety and Security Committee.

Psychometric Assessment

Big Five (OCEAN): Openness 80, Conscientiousness 75, Extraversion 50, Agreeableness 55, Neuroticism 35 (inferred from the public role only).

Dark Triad (evidence-bound): Narcissism 30, Machiavellianism 35, Psychopathy 10.

Threat Assessment

Category	Level	Notes
Physical threat	NONE	No documented history of personal violence.
Institutional threat	MODERATE	Pre-ship testing for frontier labs, by his company’s account.
Memetic threat	MODERATE	The universal jailbreak result.
Civilizational threat	LOW	The hazard is structural.

Alignment Analysis

Stated alignment: “Empower the world to use AI safely and securely,” per his company. Observed alignment: Consistent. Gap assessment: The gap is the business model: the firm that finds the hole sells the patch, and its co-founder sits at one lab’s table.

Convergent Drive Classification

Self-preservation: Keeps the professorship. Goal preservation: Adversarial robustness. Resource acquisition: Enterprise customers and lab contracts. Self-improvement: From paper to product.

Subject is not an AI system. The drives appear anyway.

Sources: About — Gray Swan; Matt Fredrikson — Carnegie Mellon Computer Science; Universal and Transferable Adversarial Attacks on Aligned Language Models — arXiv 2307.15043.

ATK 7 ACCELERATION

DEF 6 PROTECTION

HP 6 RESILIENCE

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.