Overview
- Microsoft, Google and xAI agreed to give the Commerce Department’s AI testing center early access to their models, a deal CAISI announced Tuesday to run national security evaluations before public launch.
- CAISI said it will probe stripped‑guardrail versions for hacking and misuse risks, building on earlier testing arrangements renegotiated with OpenAI and Anthropic to match the Trump administration’s directives.
- The Pentagon expanded classified‑network agreements with eight vendors excluding Anthropic, turning to firms such as OpenAI, Google, Microsoft, Nvidia, Amazon Web Services, Oracle, SpaceX and Reflection after labeling Anthropic a supply‑chain risk.
- Anthropic, which has limited its high‑powered Mythos tool to about 40–50 vetted partners under Project Glasswing, says Mythos has surfaced thousands of flaws including a 27‑year OpenBSD bug, a long‑hidden FFmpeg issue and 181 working Firefox attacks in testing.
- The White House is weighing a working group or executive order to formalize pre‑release oversight, while Anthropic pursues a reported $1.5 billion venture with Blackstone, Goldman Sachs and Hellman & Friedman to embed Claude in private‑equity portfolio companies, and it investigates reports of unauthorized Mythos access.