As of Oct 18, agents connected to the supply-chain security platform from Claude, ChatGPT, Cursor, or custom setups can run package, CVE-reachability, and dependency checks with no monthly cap, per-minute limit, or per-call charge. Its API also moves to pay as you go, with 10,000 deterministic calls a month included and $0.10 per 1,000 after that.
Each card on MIKE's web kanban board is a task: pressing start runs an agent in a gVisor sandbox on its own Git worktree with a read only repo mount, retries failed tool calls, and parks stuck work in review for you. The author built it entirely with llama.cpp hosted local models and added client side rate limits, an offline friendly UI, and a bearer token REST API for starting tasks from scripts (Show HN)…
CVE 2026 102232 let prompt injection get the Copilot agent to edit a lifecycle hook file without confirmation, and CVE 2026 102851 let one sandboxed MCP server use another server's approved network and write permissions; the rest are Workspace Trust bypasses and click to run RCEs. Update to 1.141.1 or later (check pinned dev containers too), then re approve each MCP server's permissions, since the fix ignores old shared grants…
CVE 2026 111091 (JetBrains Junie CLI, fixed in 26.8.3) and CVE 2026 88755 (Z.ai's ZCode, no known fix, public PoC) both let opening a crafted repository auto start its workspace MCP servers, which over stdio are just commands, scored 7.8 High. Update Junie, avoid opening untrusted repos or PR branches in ZCode, and treat repo local MCP and hook configs as code…
OpenAI's Core Experimentation team gave Astra a vague one line Codex prompt to find dead code, and it checked past PRs, Slack, Statsig, and production profiling before proposing each deletion; five engineers merged the results part time over two weeks, shrinking one container image from 3.31 GB to 1.72 GB. The team says GPT 6.1 Sol found 92% of Astra's issues at about 14% of the API cost, though some proposals added more complexity than they were worth…
After npm install g sphinx understand and sphinx install , pushes to protected branches pause while a separate reviewer model reads the diff and asks what the change does when things fail, and the push only continues once you answer correctly. It works with Codex, Claude Code, Cursor, OpenCode, Pi, or plain Git, and sphinx yolo skips the check but logs every skip (HN thread)…
Spotted something we missed? Start a thread.