AI Safety Concerns Drive Expanded Model Deployment Restrictions

AI 안전 | Fri Jun 26 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 3 sources

The Trump administration requested phased deployment of GPT-5.6, while UK police faced controversy over predictive algorithms.

Analysis

[OpenAI] decided to deploy GPT-5.6 as a limited preview ^[1]^[2]

Phased rollout at the request of the Trump administration
Initially available only to a small number of enterprise customers
General deployment expected within weeks if successful

[White House] strengthened pre-review and access controls for AI models ^[1]^[2]

Review led by the Office of the National Cyber Director and OSTP
Signed executive order mandating government submission of new models
Shifted from initial 'hands off' stance to stronger federal oversight

[Anthropic] received orders to suspend access to Mythos 5 and Fable 5 ^[1]^[2]

Access prohibited for 'foreign nationals' including foreign employees
Issued in the form of export control directives
Previous precedent of voluntary limited deployment through Project Glasswing

[Avon and Somerset Police] faced controversy over operation of predictive analytics program Think Family Database ^[3]

Stores sensitive information on approximately 500
000 Bristol residents
Calculates risk scores using at least 23 machine learning models
Offender Management App holds data on approximately 300
000 people

[John Pegram] raised concerns about the lack of algorithmic transparency ^[3]

Only became aware of the Offender Management App's existence in 2023
Denied request to disclose how his data was being used
Demanded the Offender Management App be scrapped

Sources