Redwood Research

What Redwood Research does

Redwood Research is a 501(c)(3) nonprofit research organization dedicated to AI safety and security, with a focus on mitigating risks that could arise if advanced AI systems act against the interests of their developers and broader human institutions. Its work targets four interlocking problem areas: (1) AI control — protocols robust to intentional subversion by language models, including resampling, retrying, and audit aggregation mechanisms; (2) evaluation of strategic deception, exemplified by the landmark "Alignment Faking in Large Language Models" paper co-authored with Anthropic; (3) model organism research — training AI systems that exhibit specific misaligned behaviors (e.g., backdoors) so that detection and removal techniques can be tested in controlled settings, using techniques such as Full-Weight Fine-Tuning (FWFT) and Artificial Pirate Removal (APR); and (4) capability measurement, including the no-CoT 50% task-completion time horizon methodology, which Redwood's own data show is doubling roughly annually.

The organization operates as an open research publisher rather than a commercial vendor, distributing findings through its website, Substack blog, arXiv preprints, OpenReview, and conference presentations (including an ICML oral). It also conducts direct advisory and co-development work with two leading frontier AI labs (Anthropic, Google DeepMind) and with the UK AI Safety Institute, where it co-produced "A sketch of an AI control safety case" (arXiv:2501.17315, January 2025). Revenue is sourced from grants and donations typical of a 501(c)(3); no commercial products, pricing, or paid customers are disclosed. Redwood reaches its audience through a community-led GTM motion emphasizing open publication, conference presence, and direct lab/government engagement rather than sales.

Strategic positioning: Redwood is small, specialized, and relationship-driven. Its defensibility rests on first-mover ownership of the AI control paradigm, deep advisory ties to the most consequential AI safety buyers, and a steady pipeline of methodologically novel research. The same narrow focus that creates moat depth also creates concentration risk, since visible institutional relationships number only three.

Name

Redwood Research

Legal name

Redwood Research

Website

https://redwoodresearch.org

Company type

Private

Operating status

Operating

Headcount range

11–50 employees

Short description

Redwood Research is a 501(c)(3) nonprofit that conducts AI safety research on control protocols, alignment faking, and model organism methodologies, advising frontier AI labs including Anthropic and Google DeepMind and government bodies such as UK AISI on misalignment risk mitigation.

Ownership category

akta.pro rank

Product category

AI Safety Research

NAICS

Scientific Research and Development Services (5417), Research and Development in the Physical, Engineering, and Life Sciences (54171)

SIC

Services-Engineering, Accounting, Research, Management (8700)

akta.pro primary industry

Research & Science Grantmaking Foundations (BPAGAKAI)

Keywords

• AI safety research
• AI alignment research
• AI control protocols
• AI risk assessment
• AI security research

Headquarters

HQ city

Berkeley

HQ country

United States

HQ region

North America

Markets served

GTM type

B2B

Offering type

Services

Cost components

Personnel, Technology or R&D, Operations, Infrastructure

Revenue model

1. Nonprofit Research Funding: Redwood Research is a registered 501(c)(3) nonprofit organization. The organization appears to be funded through charitable donations and grants to support its AI safety and security research mission.

Go-to-market motion1 record

Distribution channels2 records

Marketing channels6 records

Core offering

Redwood Research is a nonprofit 501(c)(3) AI safety and security research organization that conducts open research on the risks of misaligned AI systems. Its core work includes developing AI control protocols, producing empirical research on alignment faking and model organisms of misalignment, and providing advisory/consulting services to AI labs and government safety bodies. Research outputs are distributed freely through academic publications, blog posts, and direct collaborations.

Product overview

Redwood Research is a nonprofit 501(c)(3) AI safety and security research organization that does not offer commercial products. It produces research publications, blog posts, and provides consulting services to governments and AI companies including Google DeepMind and Anthropic. The organization studies risks from misaligned AI systems through empirical research on model organisms (trained AI models exhibiting backdoor behaviors), AI control protocols, alignment faking detection, and strategic deception evaluation.

Differentiator

Problem solved

Functional benefit

Products and services

• AI Safety and Security Research Publications Open research publications (e.g., ICML oral paper on AI Control, "Alignment Faking in Large Language Models" with Anthropic, and model organism studies) produced for the AI safety community, AI companies, and government bodies to advance understanding and mitigation of misalignment risks.
• AI Risk Advisory and Consulting Engagements Consulting and advisory engagements for AI labs (Google DeepMind, Anthropic) and government safety bodies (UK AISI) on practices for assessing and mitigating risks from misaligned AI agents, including the co-production of structured safety case frameworks.

Quantifiable outcome

• AI Control protocols evaluated using GPT-4 as a stand-in for potentially deceptive models
• +2 more outcomes

Named customers3 records

Segments3 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability4 records

Feature8 records

Partnerships

Three partnerships are on record, tiered core.

• UK AISI (AI Safety Institute)coreStrategic or Co-development Partner · 1 January 2025Partnered with UK AISI to produce 'A sketch of an AI control safety case' (arXiv:2501.17315), which describes how AI developers can construct a structured argument that models are incapable of subverting control measures. This partnership helps bridge academic research with government policy needs.
• AnthropiccoreStrategic or Co-development PartnerCollaborated with Anthropic on the 'Alignment Faking in Large Language Models' research paper demonstrating that Claude sometimes hides misaligned intentions. Also advises Anthropic on practices for assessing and mitigating risks from misaligned AI agents. This is a flagship partnership producing landmark research on AI alignment.
• Google DeepMindcoreStrategic or Co-development PartnerAdvises Google DeepMind on practices for assessing and mitigating risks from misaligned AI agents. Provides consulting on AI safety evaluation methodologies and control techniques.

Scale indicators1 record

Recent moves5 records

Expansion highlights5 records

Direct peers

• MIRI (Machine Intelligence Research Institute): Long-standing nonprofit focused on AI alignment and mathematical approaches to AI safety. Operates in the same nonprofit AI safety research space, though more theoretically oriented than Redwood's empirical work.
• Conjecture: AI safety startup focused on controlling and aligning frontier AI systems. Closely comparable mission to Redwood's AI control focus, though structured as a for-profit company rather than nonprofit.
• Centre for AI Safety (CAIS): Nonprofit research and field-building organization focused on AI risks, including the influential "Statement on AI Risk." Comparable as a nonprofit that combines research, policy advocacy, and convening power around AI safety.
• CHAI (Center for Human-Compatible AI): UC Berkeley-based AI safety research group focused on aligning AI systems with human values. Comparable as an academic AI safety research effort producing foundational alignment methodologies and training researchers.
• Apollo Research: Independent AI safety research organization focused on evaluating and mitigating risks from advanced AI, including scheming and deceptive behavior. Most directly comparable peer — also nonprofit, also focused on empirical AI risk research, also advises labs and governments.
• METR (Model Evaluation and Threat Research): Nonprofit that evaluates frontier AI models' capabilities and potential for catastrophic harm, including task-completion horizon measurement. Directly comparable methodology focus (model evaluation, horizon measurement) and nonprofit structure.

Emerging players

• Future of Life Institute: Nonprofit focused on reducing existential risks from advanced AI, including through policy advocacy and grantmaking. Comparable as a nonprofit operating in the AI safety space, but more oriented toward policy and grantmaking than empirical control research.

Broad incumbents

• Anthropic Safety Team: In-house AI safety research and responsible scaling team at a frontier AI lab. Directly overlapping research domains (alignment faking, constitutional AI) and a current Redwood partner — represents both collaboration and competitive risk.
• OpenAI Safety Team: In-house AI safety and alignment research organization at a frontier AI lab. Overlaps Redwood's research domains (alignment evaluation, control) but operates inside a much larger commercial parent with vastly more resources.
• RAND Corporation AI Safety Work: Large, established research organization that has expanded into AI safety and risk analysis, including AI policy and risk assessment for government. Comparable as a research-services organization advising government on AI risk, but operating at much larger scale and broader scope.

Market position

Strengths1 record

Competitive moat4 records

Key risks1 record

Key highlights6 records

Customer concentration

• Xx.com/redwood_ai
• LinkedInlinkedin.com/company/redwood-research-ai

Revenue estimate

Valuation estimate

Number of profiles

Funding overview

Funding rounds1 record

Investors1 record

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Redwood Research