Anthropic Grants Independent Evaluators Employee‑Level Access to Frontier AI Models
What Happened — Anthropic announced that six engineers from independent AI‑safety firm Metr will receive “employee‑like” access to its unreleased models, hardware, internal tools, and training data. The partnership is intended to let Metr conduct deep, independent risk assessments and publish findings that Anthropic cannot veto, aside from limited redactions.
Why It Matters for Trust & Control Assurance
- Demonstrates a control objective around independent oversight of high‑risk AI systems, a scenario continuous‑control‑assurance programs are built to monitor and evidence.
- Requires documented third‑party risk management, access‑level segregation, and audit‑ready evidence of evaluator activities.
- Aligns with Verisq’s Vendor Risk Management capability, which provides continuous monitoring and proof of due‑diligence for external assessors.
Who Is Affected
- AI research labs and SaaS providers deploying frontier models.
- Third‑party AI‑safety evaluators and their client organizations.
Recommended Actions
- Map the “independent evaluator access” practice to your AI governance and third‑party risk controls; capture agreements, access logs, and redaction policies as audit evidence.
- Implement continuous monitoring of evaluator activities (e.g., privileged‑access logs, data‑handling records) and integrate them into your Trust Center for defensible reporting. Source: https://www.databreachtoday.com/anthropic-gives-ai-safety-evaluators-insider-access-a-33134
Technical Notes – The arrangement gives evaluators:
- Physical and network access to Anthropic’s compute clusters.
- Read‑only access to model training checkpoints, code repositories, and internal incident logs.
- API access to unreleased model endpoints for scenario testing. Source: https://www.databreachtoday.com/anthropic-gives-ai-safety-evaluators-insider-access-a-33134