These institutes do the kind of testing a single company has reasons to soft-pedal: before or after a powerful model is released, government researchers probe whether it meaningfully helps with cyberattacks, biological or chemical weapons, or evading human control, and publish methods others can reuse. A typical output is a capability assessment of a newly released model, such as the joint UK AISI and CAISI cyber-capability assessments listed on NIST’s page.
The names have shifted with politics. The UK announced on February 14, 2025 that its AI Safety Institute had become the AI Security Institute, emphasizing national security and crime. In the U.S., the counterpart at NIST has been known as CAISI, the Center for AI Standards and Innovation; as of October 24, 2026, NIST’s page for it carries a newer name, the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), while recent posts on the same page still say CAISI. Its stated role is to be industry’s main point of contact in the U.S. government for testing, to set up voluntary agreements with developers, and to run unclassified evaluations focused on demonstrable risks such as cybersecurity, biosecurity and chemical weapons.
Why it matters: these bodies supply much of the independent evidence about frontier models, and their findings feed company frontier safety frameworks and practices like red teaming. Editorially, testing is voluntary in both countries; the institutes evaluate rather than license models.
Sources
- GOV.UK: Tackling AI security risks to unleash growth and deliver Plan for Change — February 14, 2025 press release announcing that the UK AI Safety Institute became the UK AI Security Institute, with a new criminal misuse team; checked October 24, 2026.
- AI Security Institute (UK) — States its mission is to equip governments with a scientific understanding of the risks posed by advanced AI, and lists its technical research and evaluations; checked October 24, 2026.
- NIST: Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) — NIST’s page for the center, which describes it as industry’s primary U.S. government point of contact for testing, voluntary agreements with developers and unclassified evaluations focused on demonstrable risks such as cybersecurity, biosecurity and chemical weapons; recent posts still use the CAISI name, including a July 23, 2026 joint UK AISI / CAISI assessment; checked October 24, 2026.