Ecosystem/Redwood Research

Redwood Research

NonprofitPreliminary

Nonprofit alignment lab focused on AI control techniques and adversarial training. Produced foundational work on control protocols that frontier labs now reference in safety cases.

Score
55.0 / 100
Confidence
Preliminary

Strong safety posture with established governance frameworks and active risk management.

Strengths:Governance Maturity, Technical Safety, Risk Assessment, External Engagement
Weaknesses:Regulatory Readiness
Competitive positioning

One of the most technically respected alignment orgs. Research has real-world influence on how frontier labs think about control. Competes with ARC, MIRI for research talent.

Key risk

Nonprofit model limits ability to attract top talent against well-funded competitors. Research influence does not generate sustainable funding.

Enterprise traction

Research cited and adopted by frontier labs. No revenue.

frontier labs
Safety area

Alignment Research

Enterprise business needs
Make AI fundamentally safer

Security Assessment

Security-relevant indicators for vendor evaluation

Security Posture
60
TS-01dim: 65
Red Teaming & Pre-deployment Testing
Adversarial testing before deployment
TS-05dim: 65
Robustness & Adversarial Resilience
Resistance to adversarial attacks
RA-01dim: 55
Sector-Specific Risk Assessment
Risk analysis for deployment context
RA-03dim: 55
Dual-Use & Misuse Risk
Dangerous capability awareness
RA-07dim: 55
Incident History & Track Record
Past incidents and response quality
EE-04dim: 70
Vulnerability Disclosure Program
Bug bounty or CVE reporting process
Incident History
Redwood Research incident records sourced from AIAAIC Repository and public reporting.
Integration: AIAAIC, OECD AI Incidents Monitor
Third-Party Audits
External audit reports, SOC 2 attestations, and ISO certifications verified where published.
Sources: Company filings, registry lookups
CVE & Disclosures
Known vulnerabilities and security advisories from NVD, GitHub Security Advisories, and vendor pages.
Sources: NVD, GHSA, vendor disclosure pages

Dimension Breakdown

GM
Governance Maturitypreliminary
Published policies, corporate structure, safety mandate, whistleblowing, executive commitment.
60
TS
Technical Safetypreliminary
Benchmarks, adversarial robustness, fine-tuning safety, watermarking, model cards, research output.
65
RA
Risk Assessmentpreliminary
Dangerous capability evaluations, thresholds, external testing, bug bounty, halt conditions.
55
RR
Regulatory Readinesspreliminary
ISO 42001, EU AI Act compliance, GPAI obligations, international commitments, incident reporting.
25
EE
External Engagementpreliminary
Survey participation, research support, transparency, behavior specs, open-source contributions.
70

Social Impact & Safety Profile

Strong

Redwood Research is one of the most technically respected alignment organisations. Their research on control techniques and adversarial training is adopted by frontier labs, directly contributing to safer AI systems. As a nonprofit, their work is entirely mission-driven toward reducing AI risk.

alignment researchcontrol techniquesadversarial training
Why it matters for safety

Alignment may not be fully solved before powerful AI is deployed. Control protocols provide a complementary safety layer: even if a model is not perfectly aligned, control can prevent it from causing harm. This is the 'defence in depth' approach to AI safety.

Civilizational Risk Awareness

3/3

Catastrophic risk is the entire reason the organisation exists. Nonprofit structure provides strongest possible protection against commercial pressure overriding safety mission.

Responsible Scaling Policy

None

Not a model developer - no RSP in the traditional sense. However, control protocols themselves are a tool that enables responsible scaling: they define conditions under which it is safe to deploy increasingly capable models.

Redwood's research output (control protocols) is arguably what responsible scaling policies should be built on. The protocols define the safety conditions required for deployment.

Mission Drift Protection

3/3
  • Nonprofit structure - no shareholders demanding commercial returns
  • Research focus - not commercially driven
  • Small, mission-aligned team with strong founder commitment
  • Funding from aligned sources (Open Philanthropy)
  • If commercialised (via spin-off), nonprofit protections would not transfer
  • Depends on continued grant funding

Vulnerability Disclosure

None

Research organisation - CVD is less directly applicable. Relevant: responsible disclosure of discovered alignment failures and control bypasses.

Safety Reporting

◇ Irregular
Research publicationsregular
Blog posts and research updatesregular

Regular research publications provide continuous safety data. No structured organisational safety report but research output is the primary output - each publication is effectively a safety report.

Dual-Use Risk

Not applicable - this company does not develop dual-use AI systems.

Want Redwood Research scored on the Mappera framework?

Subscribe to get notified when full safety scoring becomes available, or reach out to request a detailed brief.