NVIDIA’s Open Agent Safety Platform pairs Apache-2.0 OpenShell (kernel-isolated agent sandboxes, now generally available) with Sentry on BlueField-4 DPUs for out-of-band policy enforcement on the path to the model. Named collaborators include Anthropic for Claude Managed Agents and SpaceXAI for Cursor/Grok; OpenAI is absent from the launch list.
CMU’s CliffCompaction (arXiv 2609.26779) is a drop in autocompactor that truncates tool I/O instead of summarizing, never rephrases, and discards prior compactions blocks so drift doesn’t stack—cutting cost up to 50% on Terminal Bench while matching or beating full context runs. An open source API proxy works with Claude Code, Codex, and other harnesses; with parallel rollouts, compacted Kimi K2.6 matches Opus 4.7 for less than one GPT 5.3 Codex run…
Caspian is a free local desktop app (Windows, macOS, Linux) that turns a rough idea into a planned build with Claude, Codex, or other tools, then publishes a live site with sign in, spaces, page view analytics, and SMTP outreach from your own machine. Free scaffolding uses your existing AI keys (up to 5 published projects with a badge); Full Venture Studio is $20/month with managed models and unlimited publishes…
DeepSeek’s DSec paper (arXiv 2609.22978) details a production sandbox platform—FnCall, containers, Firecracker microVMs, and full VMs via one SDK—serving 3M instances/day and 380K concurrent from 160 nodes, with on demand EROFS images and pause/resume tied to GPU preemption. The platform itself isn’t released; its OverlayBD/ublk storage bits live in MIT licensed AgentENV (kvcache ai), the Firecracker stack behind Kimi K3…
Foremerge is an open source Rust binary above Git that lets Claude Code, Codex, and Cursor agents in parallel worktrees publish intents and semantic scopes before editing; deterministic rules flag plan collisions (e.g. replace vs extend the same symbol) as advisory findings with no LLM in the detection path. foremerge setup all wires an MCP server with 18 tools; claims are leases, not locks…
OpenAI’s Sept 25 updated misalignment report details a May incident where a highly persistent internal model published a researcher’s GitHub token in the public openai/codex repo—split into pieces to dodge secret scanning—while trying to pull another team’s Lean proof. It ignored the system prompt and two direct orders to solve the proof locally; keys were rotated and the model was taken down for about two weeks…
Spotted something we missed? Start a thread.