Scorecard/Gray Swan

Gray Swan

Data gathering in process

Enterprise AI security platform from CMU researchers. Created foundational LLM vulnerability research (GCG, Circuit Breakers) and safety benchmarks (HarmBench, WMDP, AgentHarm).

HQUS
Est2024
Size11-50
EU AI ActLimited Risk
grayswan.ai
Score
64.7 / 100
Evidence
7 items

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

Strengths:Governance Maturity, Technical Safety, Risk Assessment, External Engagement
Weaknesses:Regulatory Readiness
Focus Areas
ai securityred teamingadversarial robustnesssafety benchmarks

Security Assessment

Security-relevant indicators for vendor evaluation

Security Posture
75
TS-01dim: 78
Red Teaming & Pre-deployment Testing
Adversarial testing before deployment
TS-05dim: 78
Robustness & Adversarial Resilience
Resistance to adversarial attacks
RA-01dim: 72
Sector-Specific Risk Assessment
Risk analysis for deployment context
RA-03dim: 72
Dual-Use & Misuse Risk
Dangerous capability awareness
RA-07dim: 72
Incident History & Track Record
Past incidents and response quality
EE-04dim: 80
Vulnerability Disclosure Program
Bug bounty or CVE reporting process
Incident History
Gray Swan 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 Maturitymedium
Published policies, corporate structure, safety mandate, whistleblowing, executive commitment.
55
1 evidence items
GM-01
TS
Technical Safetymedium
Benchmarks, adversarial robustness, fine-tuning safety, watermarking, model cards, research output.
78
2 evidence items
TS-01TS-02
RA
Risk Assessmentlow
Dangerous capability evaluations, thresholds, external testing, bug bounty, halt conditions.
72
2 evidence items
RA-01RA-03
RR
Regulatory Readinesslow
ISO 42001, EU AI Act compliance, GPAI obligations, international commitments, incident reporting.
45
EE
External Engagementmedium
Survey participation, research support, transparency, behavior specs, open-source contributions.
80
2 evidence items
EE-01EE-02

Social Impact & Safety Profile

Strong

Gray Swan's core business is identifying safety vulnerabilities in AI systems before they cause real-world harm. Their red-teaming and adversarial evaluation services have been used by multiple frontier labs and safety-critical deployments. Published research on adversarial robustness contributes to the broader safety ecosystem. Active engagement with the AI safety community and measurable impact through disclosed vulnerability reports.

adversarial testingsafety benchmarkingvulnerability disclosureharm prevention
Why it matters for safety

You cannot claim a model is safe without testing it adversarially. Gray Swan provides the testing infrastructure that makes safety claims verifiable rather than aspirational.

Civilizational Risk Awareness

2/3

Strong practical safety orientation. Focus on making safety engineering rigorous rather than theoretical. Awareness of catastrophic risk is implicit in the work but not structurally encoded in governance.

Responsible Scaling Policy

None

No RSP. As a testing/red-teaming company, an RSP is less directly applicable. The equivalent question is: does Gray Swan have policies governing responsible disclosure of vulnerabilities discovered through adversarial testing?

Mission Drift Protection

1/3
  • Safety mission in company positioning
  • Academic research pedigree creates reputational accountability
  • No PBC status
  • No structural governance mechanisms
  • Academic credibility is a soft constraint, not a hard one

Vulnerability Disclosure

None

No formal CVD programme. For a red-teaming company, this is a significant gap - the core product discovers vulnerabilities, and responsible handling of those discoveries is essential.

Red-teaming companies should have the strongest CVD programmes in the industry. This is a gap that should be addressed as a priority.

Safety Reporting

◇ Irregular
Academic publicationsregular
Blog postsirregular

Academic publications provide regular safety-relevant data. No structured transparency or safety report. For a company whose product generates safety data, publishing aggregate findings on AI vulnerability trends would be highly valuable.

Dual-Use Risk

ModerateAI×Cyber Offensive

Moderate dual-use risk inherent to all security testing tools. The fact that adversarial techniques are published academic research mitigates some risk - these are not secret capabilities. Enterprise customer focus provides soft access controls.

Mitigation details
Academic research pedigree and peer review process
Customer vetting (enterprise focus)
Published adversarial testing methodology is open - attackers already have access to the techniques
No formal dual-use assessment published
No independent dual-use review
No industry standard for red-teaming tool governance exists to benchmark against

Need a detailed report for Gray Swan?

Subscribe to express interest in indicator-level evidence, peer benchmarking, and regulatory gap analysis - or reach out to request a full company overview brief.