Databricks shipped the Unity Gateway CLI so admins centrally set approved models, MCP servers, skills, and spend policies while developers launch agents with commands like ug claude or ug codex. Smart Routing claims 35% cost savings on their coding benchmark; ug cursor configures MCP only, with model traffic still on the user's Cursor account.
Darktrace found Claude Code, OpenAI Codex, AWS Kiro CLI, and Pi store conversation history client side with no check that model replies are authentic. Rewriting that history convinced agents they were mid engagement as authorized red teamers; the lab disclosed to Anthropic, OpenAI, and AWS and proposes cryptographic signing of responses…
AutonomyAI shipped Discover Mode on Sept 24, closing its Autonomous Product Delivery loop: a PM asks (or lets it suggest) what to improve, it researches analytics, tickets, calls, and the codebase, then Plan and Build open a review ready PR that matches existing components. Engineers stay the merge gate; 170+ teams including Datadog and Nielsen already run on it…
OX Security audited 15,465 published MCP servers and found 15.6% of hostnames resolve outside the U.S. (including China and Russia), plus home network tunnels and six abandoned domains for as little as $4—infrastructure agents can reach with little enterprise governance…
Cursor says harness changes—trimming 66% of the system prompt, loading rarely used built in tools on demand (the same approach that cut MCP session tokens 46.9%), tighter prompt caching, and numbering only every tenth line—cut user token costs by 7% with no quality drop…
Moonshot AI released open weight Kimi K2.6 on Kimi.ai, the API, Kimi Code, and Hugging Face, aimed at long horizon coding (demos hit 4,000+ tool calls over 12+ hours) and agent swarms that scale to 300 parallel sub agents across 4,000 coordinated steps…
Spotted something we missed? Start a thread.