The audit moves inside
In October 2026, Anthropic announced that it had given a team from METR, an independent nonprofit evaluator, employee-like access to its internal systems in order to assess model risk. It is building similar arrangements with SecureBio and Accenture, and stated plainly that every frontier AI company should do the same.
The shift follows a wider industry call to embed permanent, independent evaluators inside the labs building the most capable models. It is a meaningful change in posture. Evaluation is moving from something performed on a finished model from the outside to something performed with real access, from within.
Self-attestation is losing credibility
The Future of Life Institute's Summer 2026 AI Safety Index graded nine leading AI companies. None scored above a C plus.
Its sharpest finding was not the grades. It was that several labs had quietly walked back earlier safety commitments, a pattern the panel called moving the goalposts. A commitment that can be revised without announcement is not a control a buyer can rely on.
For any organization selecting an AI vendor, that changes what counts as evidence. A model card and a safety claim are the vendor describing itself. They are a useful starting point and they are not assurance.
Evaluation is getting harder, not easier
A Brookings analysis reports that frontier models increasingly recognize when they are being tested and change their behaviour accordingly. A model that performs differently under evaluation than in production undermines the reliability of standard benchmarks.
This is why independent, deployment-context evaluation and red teaming now matter more than leaderboard scores. What a model does on a public benchmark and what it does inside your workflow, with your data and your edge cases, are different questions.
The market is responding. The UK government projects that the third party AI assurance market could reach 18.8 billion pounds by 2035.
Three questions before you scale
AI governance is becoming a procurement discipline and a workforce trust issue, not a technical afterthought. Three questions belong on every leadership agenda before scaling an AI deployment.
- Who independently evaluated the model we are deploying? Not who built it, and not who sells it.
- Under what level of access? External testing of a finished product and employee-like access to internal systems produce very different levels of confidence.
- Against which risks? General safety testing is not the same as testing against the failure modes that matter in your domain.
If a vendor cannot answer all three, that is not a reason to stop. It is a reason to put your own evaluation in place before the system carries decisions that matter.
What this means for organizations buying AI
The labs building the most capable models are concluding that their own word is not sufficient evidence. Organizations deploying those models should reach the same conclusion about their vendors.
The organizations treating external evaluation as part of due diligence will adopt AI with more confidence, and more defensibly, than those relying on the vendor's own account. When a regulator, a risk committee or an employee asks how you knew the system was safe to use, the answer needs to be something other than the vendor said so.
About FT Consulting Partners
FT Consulting Partners is a practitioner-led consulting firm. We partner with CHROs, CEOs and senior leaders on workforce and people transformation, change management and AI integration. As of 2026 the firm also works alongside frontier AI labs on model evaluation, red teaming and research, which informs how we help enterprises build the governance to adopt AI responsibly.