AI Security Institute (AISI) on AI labs should give independent evaluat…

AI Security Institute (AISI)

UK government AI safety institute

It is clear from our experience that to run high quality evaluations that elicit high fidelity information to the potential risks posed by frontier systems, it is important for independent evaluators to have: Access to a Helpful Only (HO) version of the model, alongside the Helpful, Honest and Harmless (HHH) version of the model that will be deployed, the ability to turn off/on trust and safety safeguards, and fine-tuning API access.

AI Verified (Oct 2024)

Source:

AI Security Institute, “Early lessons from evaluating frontier AI systems,” October 2024, section 7: What is needed for effective testing?

20d ago

Policy proposals and claims

AI labs should give independent evaluators employee-like access

votes For

Statement relation verification history AI Verified Report this

AI Verified The AI Security Institute directly specifies model and safeguard access needed by independent evaluators to conduct high-quality frontier-system evaluations. · Hector Perez Arenas gpt-5 · 20d ago

Vote inference verification history AI Verified Report this

AI Verified Recorded answer is for. The AI Security Institute says independent evaluators need access to model variants, safeguards, and fine-tuning to evaluate frontier risks properly. · Hector Perez Arenas gpt-5 · 20d ago

Quote authenticity verification history

Report this

AI Verified The AI Security Institute page contains the stored wording in its section on effective testing and identifies the needed access for independent evaluators. · Hector Perez Arenas gpt-5 · 20d ago

replying to AI Security Institute (AISI)

Other opinions from AI Security Institute (AISI)

See all

AI Security Institute (AISI) votes For

Open-source AI is more dangerous than closed-source AI

For models with dangerous capabilities – including highly cyber-capable models – open weight release therefore creates a persistent and irreversible risk of misuse.

AI Verified source (Jul 17, 2026)

added 2mo ago

AI Security Institute (AISI) votes For

Require AI labs to publish safety evaluations before deploying frontier models

As AI agents are given more autonomy and harder tasks, underestimating their capabilities becomes more consequential.

AI Verified source (Jul 2, 2026)

added 1mo ago

AI Security Institute (AISI) votes For

Governments should fund open-source AI safety tools and red-teaming infrastructure as public goods.

By open-sourcing not only Inspect, but also our methods and infrastructure where we can, we hope to help researchers, organisations, and other evaluators stand up rigorous evaluation capability without starting from zero.

AI Verified source (Jun 18, 2026)

added 2mo ago