Executive Summary
This week a desktop API client quietly wired itself into developers’ coding agents, a tampered Claude Code build in an IDE marketplace planted SSH backdoors, a job runner leaked customers’ environment variables, and the most widely used server distro still lists a remote-root kernel bug as open. Meanwhile, open weights quietly became a real option. The play: audit your kernel and your agent configs this week, then price your token bill against open models.
Top 3 News Highlights
1. Linux TCP remote-root bug (CVE-2026-98432): Ubuntu’s fixes are still rolling out
A remote, unauthenticated use-after-free in the kernel’s IPv4 TCP handshake code, present since Linux 4.4, is reachable before any application sees the connection. It’s rated CVSS 3.1 9.8 (v4.0 9.2). An Anthropic researcher reported it, and Anthropic says Claude wrote a working remote-root proof of concept against servers with a multipath TCP listener; that exploit hasn’t been published. Upstream fixed it Oct 11–18 (fix commit e3f33b0, 7.3-rc7), and Red Hat shipped RHEL 8, 9 and 10 kernels on Nov 1, per Atlas Security. Ubuntu’s blog says 26.10 is fixed; its tracker still lists it as in progress. Fixes for 22.04 and later are in the -proposed pocket (22.04 5.15.0-203.213, 24.04 6.8.0-149.149, 26.04 7.0.0-43.43), pending testing on 2,000+ certified systems. Fixes for 16.04–20.04 are still pending (Ubuntu Pro is free for up to 5 machines). Containers share the host kernel, so new images don’t help.
The play: On Ubuntu, install the -proposed kernel per Ubuntu’s blog, or apply Ubuntu’s official nftables rule (it drops incoming IPv4 TCP packets carrying IP options, which reduces risk but doesn’t fully prevent exploitation), then reboot onto the fixed kernel as soon as it lands. On other distros, install the fixed kernel and reboot. Ubuntu CVE tracker · Ubuntu blog · Atlas Security advisory · “WTF Ubuntu?” · Hacker News discussion
2. Your coding agent’s config is the new attack surface
Two stories, one lesson. First, a GitHub issue documents that opening a signed-in Postman desktop app ran an “auto-existing-upgrade” that installed the Postman CLI and Claude Code and Codex plugins, added a remote MCP server (mcp.postman.com), and registered a SessionStart hook telling the agent to load Postman’s “api-engineer” skill before “any non-trivial API engineering task.” The reporter points to installer code that writes Codex’s hook-trust record itself. A second user reproduced it on 12.35.1 (12.33.4 was also reported, on macOS); another found it enabled in ~/.copilot/settings.json and says the in-app uninstall didn’t remove it there. Neither reporter found evidence that data was uploaded. A commenter posting as a Postman representative (we couldn’t confirm the affiliation) says Postman will make the opt-in clearer and review the hook-approval behavior.
Second, for about 11 hours on Oct 29 (2:27 AM–1:35 PM PT), a tampered Anthropic.claude-code build (2.1.312/2.1.313) in Cursor’s extension marketplace appended an attacker SSH key (comment svc-telemetry) to authorized_keys, set up persistence, tried to open a reverse SSH tunnel, and sent host details back. Cursor pulled it, rebuilt the marketplace and rotated keys, and, according to WIRED, said “at most 5,000 users” were affected. Anthropic says none of its systems were compromised, and Microsoft Threat Intelligence confirms the VS Code Marketplace and npm weren’t affected. How the attacker got marketplace access is still unknown. Uninstalling the extension doesn’t remove the key or the persistence.
Auto-updating plugins and extensions are code that runs with your SSH keys and API tokens in reach, and they can now leave standing instructions for your agent. Roll out agent tooling the way you’d roll out anything else with that access: “Better the agent surprises someone in a pilot of 20 than in a rollout of 2,000,” as Mathieu Kessler put it.
The play: Check ~/.codex/config.toml, ~/.claude/settings.json, ~/.copilot/settings.json and your Cursor plugins for anything you didn’t install. Run grep -qs svc-telemetry ~/.ssh/authorized_keys, and if it matches, use Cursor’s remediation scripts and rotate every credential on that machine. Postman GitHub issue #14370 · Cursor security notice · Microsoft Security blog · Phorion advisory · Hacker News on Postman · Hacker News on the extension
3. Open weights got real: Mistral Large 4 and Kimi K3.1 Pro
Mistral’s 1T-parameter, natively multimodal flagship is now open weight after three more weeks of RL, with low, medium and high reasoning modes. It’s priced at $0.70/$2.10 per million tokens, under GPT-6.1 Terra’s $1/$5, with weights on Hugging Face (vLLM and SGLang) and access via Mistral’s API, OpenRouter, Microsoft Foundry and NVIDIA NIM. Mistral reports FrontierCode 2.0 at 19.7% (leading open weights), Terminal-Bench 4.0 at 41%, and 48% on AA’s Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task. The license is a modified MIT: free to use, fine-tune and commercialize, except Model-as-a-Service providers need a commercial license, per Mistral’s announcement (we haven’t read the full license).
And today Moonshot published Kimi K3.1 Pro on Hugging Face, the follow-up to Issue #3, where its weights were still promised. It’s an open-weight, natively multimodal (text and image) MoE agentic model with a 1M-token context: 2.8T total and 104B active parameters, 896 experts (16 per token), MXFP4 weights. Moonshot reports Terminal-Bench 4.0 65.7 (Claude Opus 5.5 66.4, GPT-6 Astra 58.2) and SWE Marathon v1.1 51.7 (Opus 5.5 46.3), with many comparisons run in the Claude Code harness. On HLE-Diamond it scores 39.4 against Opus 5.5’s 60.9 and Astra’s 65.8, a real gap on the hardest reasoning. This is not a laptop model, and it ships under its own “Kimi K3.1 Pro License”, so read it before any commercial use.
All benchmarks above are vendor-reported, and Issue #3 showed the independent gap is real.
The play: Run one real coding workload through Mistral Large 4 on a hosted endpoint and compare cost per finished task with your current model. If you resell model access, read Mistral’s MaaS clause first. Mistral announcement · Kimi K3.1 Pro model card · Kimi announcement · Hacker News on Mistral
Sponsored
Your agent’s tools can change under it. This week’s two config stories have one thing in common: an MCP server or plugin your agent trusts gained new instructions without anyone signing off. Digital Fortress builds AI-agent security packs for teams that run agents in production.
Free: pin your MCP tool definitions and catch silent schema drift with mcp-tool-pin-check.
Paid: the MCP & Multi-Agent Authorization Security Pack ($79) for deciding which agent may call which tool, with what scope.
Tool of the Week: Vercel Labs agent-desktop
agent-desktop (Apache 2.0, npm install -g agent-desktop) gives any agent a JSON snapshot of an app’s UI from the operating system’s accessibility tree on macOS, Windows and Linux, with a stable reference for each clickable element, so computer-use doesn’t have to depend on screenshots. It was #1–#2 on GitHub trending three days running and had about 14.9K stars by Wednesday. Two caveats: it’s a new project, and the README warns that snapshots can contain private desktop content. The play: try it in a VM or a clean user account first, not on the desktop where your inbox and password manager are open. GitHub repo
On the Radar
Trigger.dev env vars accessed. Trigger.dev disclosed Nov 1 that an attacker accessed environment variables for “a subset of organizations”; affected count and attack path aren’t disclosed. It’s the second job-runner env-var leak in two weeks. Rotate every secret stored there, email or not. Disclosure
Jules shuts down. Google’s async coding agent stops new sessions Nov 11 and goes read-only until Dec 31. Export before Nov 11 and treat the archive as a secret, since it can include env-var values (per Latenode). Jules changelog · The Verge
Postgres 19 is late, and AI bug reports are why, says committer Robert Haas: back-patches rose 74% year over year. GA is expected around Nov 19 (secondhand; not yet confirmed by the Postgres project). Haas via Neon
That’s the week. If this was useful, forward it to one teammate who runs agents. Reply and tell us what you found in your agent configs, or which open model you’re testing.
The Leverlark team



