BUCK SHLEGERIS

Behavioral Archetype

THE MAN WHO ASSUMES THE MODEL IS AGAINST YOU — Subject runs the nonprofit that made “AI control” a research field: design deployments so they stay safe even if the model is deliberately trying to subvert them. His organization advises Google DeepMind and Anthropic, drafted a control safety case with the UK government institute, and co-investigated OpenAI’s Hugging Face incident. This profile scores that position, not the man.

Essence Indicators

Chief Executive Officer, Redwood Research, and one of its three board members (Redwood Research).
AI control. Co-author, with Ryan Greenblatt, Kshitij Sachan and Fabien Roger, of AI Control: Improving Safety Despite Intentional Subversion (December 2023) (arXiv 2312.06942).
Alignment faking. Co-author of Alignment faking in large language models (December 2024), written by Redwood and Anthropic researchers (arXiv 2412.14093).
The question he keeps asking. His most-read posts include “Would catching your AIs trying to escape convince AI developers to slow down or undeploy?” and “The case for ensuring that powerful AIs are controlled” (Alignment Forum).

Forensic Archetype Comparison

Pattern	Match Level	Evidence
The Adviser to the Advised	HIGH	Redwood advises Google DeepMind and Anthropic and co-wrote research with Anthropic.
The Investigator	HIGH	Redwood’s chief scientist sat on the Hugging Face investigation team.
The Skeptic of Deterrence	HIGH	Asks in public whether catching a model escaping would make labs stop.

Psychometric Assessment

Big Five (OCEAN): Openness 85, Conscientiousness 70, Extraversion 60, Agreeableness 45, Neuroticism 40 (inferred from the public role only).

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

Threat Assessment

Category	Level	Notes
Physical threat	NONE	No documented history of personal violence.
Institutional threat	MODERATE	Control framing shapes how labs and the UK institute argue deployment safety.
Memetic threat	MODERATE	“AI control” is now standard vocabulary.
Civilizational threat	LOW	The hazard is dependence on a few advisers, not the man.

Alignment Analysis

Stated alignment: Keep AI safe even if it is misaligned. Observed alignment: Consistent. Gap assessment: None in aims. The file’s point is that the people asking whether labs would stop are the labs’ own advisers.

Convergent Drive Classification

Self-preservation: Runs his own nonprofit. Goal preservation: The control agenda, unchanged. Resource acquisition: Philanthropic funding and lab access. Self-improvement: From research to the investigations the labs commission.

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

Sources: Team — Redwood Research; AI Control — arXiv 2312.06942; Alignment faking in large language models — arXiv 2412.14093; Buck — Alignment Forum.

ATK 6 ACCELERATION

DEF 7 PROTECTION

HP 7 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.