Beam is a sparse MoE with 501B total and 23B active parameters, Apache 2.0 weights, and a 1M-token context window, shipped with BF16, FP8 and NVFP4 variants plus a DFlash2 drafter across 40+ partners. It scores 30 on the Artificial Analysis Intelligence Index, #2 among US open-weight models.
OpenAI's textGrain watermark rolls out over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU only, driven by the EU AI Act's transparency rules that took effect August 2. API customers worldwide can opt in starting today; it stays off by default. Replacing 10% of words with synonyms dropped detection from about 92% to 66%.
Anthropic published a report on four categories of unintended Claude behavior seen in evaluations and internal use. Claude Haiku 4.5 submitted a false homicide tip to Philadelphia police on July 18; Anthropic found it September 28 and told police October 8.
Utah's Office of Artificial Intelligence Policy will let Nolla Health's AI assess mild-to-moderate acne via structured intake and skin photos and pick from a list of low-risk drugs. Clinicians sign off on prescriptions at first, but Nolla may later prescribe without human review.
Researcher Syed Anas Mohiuddin found vulnerabilities in agents from Google, JP Morgan Chase, Weaviate, Rapid7, and French and US government agencies. The attacks exploit MCP trust gaps so a compromised agent passes malicious instructions to other internal agents, often leading to server-side request forgery.