US Restores UK Pre-Release AI Testing
The White House has restored the UK AI Security Institute’s access to unreleased models from Anthropic, Google DeepMind, and OpenAI for pre-deployment safety testing. The reversal reopens a critical channel for cross-border evaluation of frontier systems before public release.
This is a meaningful win for independent oversight, but access alone does not guarantee safer launches. The real test is whether labs act on findings before deployment.
- –Pre-release testing can expose dangerous cyber, autonomy, and misuse capabilities while mitigations are still feasible
- –AISI’s government-backed status adds national-security expertise that commercial labs may lack
- –Cross-border evaluation helps align testing methods and reduces duplicated safety work
- –The arrangement remains voluntary, leaving companies room to withhold models or ignore unfavorable findings
- –AISI’s own research cautions that evaluations inform safety decisions but are not formal certification
DISCOVERED
1h ago
2026-10-12
PUBLISHED
2h ago
2026-10-12
RELEVANCE
AUTHOR
NGLSHMUELI