Disclosed Oct 26, prompts over 681,280 tokens to @cf/zai-org/glm-5.3 between Aug 27 and Sep 9 hit an SGLang disaggregated prefill/decode bug that left stale KV cache pages in place, so replies could include text from other tenants' requests. Cloudflare says customers need do nothing, but anyone self-hosting SGLang in disaggregated mode should upgrade to v0.5.22 or later.
Since v2.1.311, setting CLAUDE CODE EVEN MORE DANGEROUSLY SKIP PERMISSIONS=1 lifts the short blocklist (like rm rf ) that dangerously skip permissions still prompts for, and Anthropic's Daisy Hollman confirmed it's not a joke. Managed policy settings still apply on top, so if you run unattended agents, keep limits enforced outside the prompt…
Released Oct 27, you can now route Codex through Google Cloud with model provider = "gcp" , GPT 6.1 Terra joins the model catalog, and fresh profiles default to memory v2, which can import other agents' memories. It also closes a marketplace URL check gap that could let Git connect to a host other than the one allowed, so it's worth updating…
In public preview since Oct 27, you give a GitHub App the read only Copilot Requests org permission, install it, and pass its short lived installation token to the Copilot SDK, so your agent's usage is billed to the org under its model policies and budgets. Owners can spot app usage by app client id in AI usage CSV reports and cut access by suspending or uninstalling the app…
Announced Oct 27 and available Oct 28 from $20 per user, SafeWrite intercepts Bob's file edits and package installs in memory, swapping known vulnerable versions for the nearest fixed one, blocking installs with no fix and refusing direct lockfile edits. It also adds real time SAST and supply chain scanning plus read only attack surface and file sweep scans, with /review on pull requests coming in November for GitHub…
Released Oct 27 from 100 million September PRs across 1,300+ companies, Jellyfish's report puts Claude Code at 31% of AI merged PRs versus GitHub Copilot's 23%, with agents involved in 28% of merged PRs at the median company, up from about 2% in January. Yet the median company merges just 0.5% of PRs on AI review alone, and median Claude Code spend hit $63 per developer a week (DevOps.com)…
Spotted something we missed? Start a thread.