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.
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…
Prismor (formerly Immunity Agent) is a self hosted Apache 2.0 control plane that sits between Claude Code, Codex, Cursor, or Windsurf and every tool call, allowing, warning, or blocking by policy before execution. It wraps npm/pip with supply chain scoring, cloaks secrets in payloads, and ships an MCP Gateway that policy checks calls and injection scans responses; pip install prismor plus prismor setup wires named postures from lightweight dev safe to airgapped…
MiniMax turned on M3.1 Flash Preview inside MiniMax Code on Sept 27 with no model card, public API listing, or pricing—just a 1M token context and five reasoning effort tiers (low through max). Days earlier, community fingerprinting linked OpenRouter’s free stealth/space bunny alpha preview to the same MiniMax family; MiniMax has not confirmed the match…
YC W26’s /dev/fast open sourced Whiteboard, a local desktop canvas where Claude Code, Codex, and similar agents draw sequence and ER diagrams of their work next to an AST aware Rust diff viewer and Decision Log. Diagram nodes jump into a bundled Code OSS editor; installs cover macOS, Linux, and Windows at install.dev.fast ( 1.9k GitHub stars)…
Hayder Tirmazi’s Sept 26 write up speeds llama.cpp’s n gram prompt lookup drafting up to 42× and cuts peak memory 2.6× via reference reads, dense outer maps, sorted vector followers, and a Lemire constmap for the static cache—acceptance rate unchanged. Gains concentrate on repetitive drafting (code edits, structured output), not general chat…
Spotted something we missed? Start a thread.
