Appeals court upholds Pentagon's ban on Claude over its safeguards

The 2-1 ruling splits from a California court that called the ban illegal, setting up a likely Supreme Court fight — as Codex gains Bedrock support.

Nowline SEP 26 9:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Anthropic chose the exit over the edit

    In June, Defense Secretary Pete Hegseth banned Claude from Pentagon systems and defense contractors after Anthropic refused to swap its usage policy — which bars mass surveillance and autonomous-weapons work — for blanket “all lawful uses” language. On Sept 25 a D.C. Circuit panel upheld the ban 2-1 under FASCSA, finding the Pentagon had “ample support” to call Claude a supply-chain risk.

  • A split now headed for a higher court

    The decision directly contradicts a California federal court that struck the same ban down last month, leaving conflicting precedent on whether the government can blacklist an AI vendor for keeping its own guardrails. Anthropic says it “respectfully disagrees” and can seek en banc rehearing or the Supreme Court — the result will shape how every lab answers pressure to loosen model policies.

  • Codex can now run GPT-6 through your own Bedrock

    Codex 0.157.0 (Sept 25) adds GPT-6 Sol and Luna with Amazon Bedrock support, so you can route Codex through your existing AWS account instead of hitting OpenAI directly. The release also lifts the file-upload timeout from 60s to 5 minutes with retries and turns on fullscreen transcripts by default.

  • Elsewhere: GPT-6 Astra reads your assembly mistakes

    OpenAI's GPT-6 Astra scored 80% on a furniture-assembly benchmark — flagging where a photographed IKEA build went wrong — versus 28% for late-2025 models, and ahead of Claude Fable 5.1 (70%) and Opus 5 (61%). At ~3 minutes per photo it's too slow for live help, but it's a real signal for visual-inspection and troubleshooting apps.