Story perspectives

CAISI Secures DeepMind, Microsoft, xAI for Security-Focused AI Testing

5/6/2026

Summary

CAISI Secures Testing

On May 5, 2026 Center for AI Standards and Innovation (CAISI) announced that Google DeepMind, Microsoft and xAI will provide AI models for pre-release testing.
The firms signed agreements to submit models for evaluation of frontier AI capabilities and security risks.
The agreements let CAISI test models for cybersecurity, biosecurity and chemical-weapon risks before public release, while the White House is considering an executive order for a pre-release review alongside the deals.

Summary

CAISI Enables Feedback

The CAISI-convened TRAINS Taskforce will receive regular feedback from evaluators during model assessments.
Evaluations will be conducted in classified environments so government analysts can examine capabilities without public exposure.
The agreements aim to facilitate information-sharing that can drive voluntary product improvements across the AI industry.
Secretary Howard Lutnick emphasizes rapid flexibility in the agreements to keep pace with fast-moving AI advancements.

Summary

Pentagon Secures AI

Microsoft’s shares slipped 0.6 percent in midday trading after the announcement.
Alphabet rose 1.3 percent on the same day.
The Pentagon secured AI access from seven firms—Google, Microsoft, Amazon Web Services, Nvidia, OpenAI, Reflection and SpaceX—broadening its provider base.
The defense pact is intended to augment warfighter decision-making in complex operational environments.
Anthropic was left out of the Pentagon agreement after a public dispute over military-use guardrails.

History

Historical Context

In February 2019 OpenAI withheld the full 1.5-billion-parameter GPT-2 model, citing misuse risks. The decision followed its 2015 nonprofit founding and a 2020 exclusive licensing deal with Microsoft that deepened corporate-government ties. Now the CAISI government has signed evaluation agreements with Google DeepMind, Microsoft and xAI, extending the 2019 caution into formal pre-release oversight.

1 of 4

CAISI Secures Testing

On May 5, 2026 Center for AI Standards and Innovation (CAISI) announced that Google DeepMind, Microsoft and xAI will provide AI models for pre-release testing.
The firms signed agreements to submit models for evaluation of frontier AI capabilities and security risks.
The agreements let CAISI test models for cybersecurity, biosecurity and chemical-weapon risks before public release, while the White House is considering an executive order for a pre-release review alongside the deals.

1 / 4