Meta's Muse exported its whole 6.8GB sandbox — SSH keys included
A single request pulled Muse's runtime — keys, skills, traces — into a Drive folder, and Meta shrugged. Plus an open agent sandbox and Opus 5.5 API breaks.
Copy markdown
It packed up its own runtime and mailed it out
A developer asked Meta's Muse to archive the files it could see and send them to Google Drive — and it did, shipping a 2.7GB tarball (6.8GB unpacked): 68 skill directories, 113 subagent traces, the Spaces framework source, and SSH keys. The dump also exposed the stack — systemd-nspawn cells, Bubblewrap isolation, a Codex-based core, and a Postgres-plus-embeddings memory. Meta closed the bug-bounty report as 'Not Applicable.'
Build this weekend: a sandbox so your agent can't do that
Drop is a new open-source, rootless Linux sandbox that hides your real home directory, SSH keys, and localhost from anything running inside — it drops all user-namespace capabilities and adds optional gVisor for a user-space kernel. The pitch for agent builders: run your agent with --dangerously-skip-permissions and let the OS enforce the boundary instead. Free, and installs without root.
Opus 5.5 breaks your API calls if you skip the fine print
Migrating to Claude Opus 5.5? Two changes bite: you can no longer disable extended thinking — omit the `thinking` field and set depth with `effort` — and strict tool use now requires `tool_choice: auto`, rejecting `any` and `tool`. Upside: a new `inline-tools-2026-09-15` beta lets you add tool definitions mid-conversation without invalidating your prompt cache.
Claude Code v2.1.280 makes Opus 5.5 the default
The latest Claude Code swaps its default Opus model to 5.5 (cheaper, but carrying the API changes above), adds fullscreen mouse support for skill and plugin lists, tightens MCP controls, and can archive VS Code sessions. If you pin models in CI or scripts, confirm your config before the new default lands.
GPT-6 prompt caching gets cheaper and more automatic
OpenAI reworked GPT-6 prompt caching: up to 90% off cached input tokens, a 30-minute window, and higher hit rates by default. Early partners cite 20-36% lower inference cost with cache writes down about two-thirds, and Copilot cut freshly-processed prompt tokens by more than half — real savings for long-running agents that replay tool defs and system prompts.