Particle.news

Anthropic Scales Back Safety Pledge as Pentagon Issues Ultimatum Over AI Use

The company replaced its binding pause-on-training policy with a flexible framework under intensifying defense demands.

Overview

  • Defense Secretary Pete Hegseth threatened blacklist measures and possible Defense Production Act action, setting a Friday deadline for compliance.
  • CEO Dario Amodei maintained that Claude should not enable autonomous weapons or mass surveillance of Americans.
  • Anthropic confirmed a shift to a non‑binding safety roadmap, dropping a requirement to halt model training when controls lag.
  • A report by Gambit Security alleged an attacker used Claude and other models to exfiltrate roughly 150 GB from Mexican systems, exposing about 195 million records.
  • Anthropic said it disrupted the activity and blocked implicated accounts, while some Mexican agencies denied breaches, a dispute playing out as the firm holds access to classified Pentagon networks and a contract near $200 million.