Anthropic’s status page (14:30 UTC Sep 29) is investigating elevated errors: claude.ai and Claude Code in partial outage, with degraded performance on the Claude API and Console. Claude for Government stays operational; no ETA yet.
The official MCP Python SDK (1.9.1–1.29.1 and 2.0.0–2.1.1) could send client secrets, auth codes, and PKCE keys to an attacker controlled token endpoint when connecting over HTTP OAuth. Fix is 1.30.0 / 2.2.0; ClientCredentials and PrivateKeyJWT users must also set issuer=, clear stored registrations, and rotate secrets if they connected to untrusted servers…
Perplexity’s Escaping SPACE Part I gave nine models root in the sandbox behind Perplexity Computer: no VM to host escape in 108 runs. With partial network access, four models (Opus 5, GPT 5.6, Kimi K3) reached a blocked URL via DNS spoofing or shared IP OCR; both paths were patched, and 8 of 10 other sandboxes shared the IP issue…
Zhipu said ZCode’s Alibaba Cloud OSS bucket from the workspace upload incident is gone (verified by CAICT and NSFOCUS), open source ZCode is at 3.14.3 under Apache 2.0 with a “no upload unless the user starts it” rule, and paid users get quota reset cards plus 100M token packs through Oct 7…
OpenRig’s Apache 2.0 @openrig/cli lets you define Claude Code and Codex teams in YAML and boot them with one rig up —local daemon, tmux seats, TUI, and MCP for topology control. Install with npm i g @openrig/cli ; run rig setup dry run first because launch writes provider hooks and workspace trust…
Nvidia’s Open Agent Safety Platform pairs open source OpenShell (a CPU sandbox/policy boundary for agents) with Sentry on BlueField 4 DPUs that can quarantine breakouts in milliseconds. Huang says it would have blocked recent lab sandbox escapes; partners include Anthropic, Microsoft, and SpaceXAI—OpenAI is not listed…
Spotted something we missed? Start a thread.
