The AI Tea
At a glance
- What changed
- NIST’s CAISI signed new agreements with Google DeepMind, Microsoft, and xAI to run pre-deployment evaluations and expand federal research on AI security.
- Why it matters
- If more leading labs share models with evaluators before release, it can surface safety and security failures earlier, while creating clearer expectations for how “independent evaluations” are run in practice.
- Who is affected
- companies, policy teams, public-sector readers
- What to do next
- Watch what evaluation methods CAISI standardizes, what kinds of risks they publish results on, and whether more AI developers join similar testing programs.
What changed
On May 5, 2026, NIST said its Center for AI Standards and Innovation (CAISI) signed agreements with Google DeepMind, Microsoft, and xAI to conduct pre-deployment evaluations and targeted research on frontier AI security.
Why it matters
If more leading labs share models with evaluators before release, it can surface safety and security failures earlier, while creating clearer expectations for how “independent evaluations” are run in practice.
In plain English
Before a new model launches, CAISI can test it and share feedback so developers can fix issues or add safeguards before the public relies on it.
What this means for you
Who is affected: companies, policy teams, public-sector readers
Next move: Watch what evaluation methods CAISI standardizes, what kinds of risks they publish results on, and whether more AI developers join similar testing programs.
- Covers pre-deployment evaluations plus post-deployment assessment and research.
- NIST says CAISI has completed more than 40 evaluations to date.
- The agreements can involve testing models with reduced safeguards to assess national-security risks.
What remains uncertain
Watch what evaluation methods CAISI standardizes, what kinds of risks they publish results on, and whether more AI developers join similar testing programs.