Michael Karvan

Nice!

Thanks to @trevin for amazing @illo_skill

AgentShield kept growing. Since the last update, it’s moved from “scan agent installs” toward real enforcement across agent runtimes. Here’s what’s new: 🧰 System package protection — detects apt, brew, yum/dnf, apk, pacman, zypper, snap, flatpak installs. Agents can’t quietly jump from pip install to apt-get install. 🧪 System package CVE scanning — opt-in live CVE checks for OS/distro packages via OSV + distro trackers, with severity floors and finding caps. 🔒 Fail-closed enforcement — shell expansion, VCS URLs, remote requirements files, unsupported managers, and untrusted conda channels get blocked instead of guessed. 🧬 Wider package-manager coverage — pip, pipx, uv, poetry, conda, npm, yarn, pnpm, bun, cargo, gem, go, and system package managers are all recognized through one shared parser. 🧱 Proxy enforcement — route pip/npm traffic through an AgentShield index proxy so resolved packages and transitive deps are scanned before they’re served. 🧷 PATH shim + exec interception — catches normal package-manager calls, PATH resets, and absolute-path installs like /usr/bin/pip install. 🤖 Claude Code + Codex hooks — PreToolUse integration scans Bash commands before agents execute them. 🪽 Hermes plugin — real pre_tool_call guard, installed into Hermes’ own runtime, blocking unsafe shell installs before tools run. 🦀 OpenClaw plugin — real before_tool_call Node plugin. Blocks unsafe exec installs instead of relying on a prompt-only skill. 📡 Daemon modes — MCP, Unix socket IPC, and local HTTP API for integrating AgentShield into agents, CI, and scripts. 🧾 SBOM + posture endpoints — MCP/API support for CycloneDX SBOMs and security posture reports. 1,000+ tests. 3 language ecosystems. System packages too. Fully local, no telemetry. pip install git+github.com/mkarvan/AgentS… GitHub: github.com/mkarvan/AgentS… #AIAgents #CyberSecurity #SupplyChainSecurity #OpenSource

Super cool

@petergyang @mvanhorn The biggest takeaway is “adding value” it doesn’t matter who you are, what degree you have. Just add value.

My top 5 takeaways from @mvanhorn: 1. Stop clicking around websites. Print them into CLIs your agents can run. Matt built Printing Press to turn websites and apps into command-line tools for agents. It researches docs, finds niche repos, and uncovers hidden API calls so agents can use sites like Google Flights, Suno, and OpenArt. Printing Press: github.com/mvanhorn/cli-p… 2. Use Compound Engineering to build. Matt says his favorite way to build is to run /ce-plan and /ce-work over and over. This system from @every gives agents structure for planning, coding, reviewing, and shipping so you are managing the work instead of writing every line yourself. Compound Engineering: github.com/EveryInc/compo… 3. The plans are for the agent, not for you. Matt often does not read AI’s full plans. Instead, he asks targeted questions: What files will you touch? What is most likely to break? What would an expert say is wrong with this idea? That catches more issues than skimming a giant plan. 4. Use last30days to do research. Matt’s last30days skill hit #1 trending on GitHub because it answers a real problem: AI research is often stale or generic. last30days scans Reddit, X, YouTube, Hacker News, Polymarket, and the web to show what people are saying right now. last30days: github.com/mvanhorn/last3… 5. To become an AI builder, solve your own problems and keep shipping. Matt has no CS degree and says he does not know how to read code. But he kept building tools for his own problems, contributing to open source, and launching what helped him. Matt walked through Printing Press, Compound Engineering, last30days, Agent Cookie, and his full agentic engineering workflow in our episode. 📌 Watch the full episode now: youtu.be/BxEf3RqIHkw

@opencode That’s awesome 👏 any chances we get GLM-5.2 soon?

Kimi 2.7 Code now available in Go text · image · optimized for coding similar pricing as 2.6

@Kimi_Moonshot can't wait to try it!

🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less overthinking, with 30% lower reasoning-token usage compared to K2.6. 🔷 Long-horizon coding: Improved instruction following, higher end-to-end coding task success rates. ⚡️ 6x High-Speed Mode coming soon! 🔌 Available today via Kimi API and Kimi Code. 🔗 Kimi Code: kimi.com/code 🔗 API: platform.moonshot.ai

AgentShield has grown from a CVE scanner into a full agent security platform. Here’s what’s new: 🔍 License compliance — auto-deny GPL/AGPL packages with configurable policy. Your legal team will thank you. 📊 Trust scores — every package gets a 0-100 reputation score based on download count, age, maintainer count, and scan history. New package with 12 downloads? Flagged. 🐳 Docker scanning — agentshield scan-docker Dockerfile parses every RUN pip/npm/cargo install and scans them all. ⚡ Drift detection — tracks scan history. If a package you approved last week just got a critical CVE, AgentShield catches it on the next run. 🚦 Agent rate limits — cap installs at N packages/hour per session. A compromised agent can’t spam-install 100 packages. 🔀 Diff scanning — agentshield diff-scan old.txt new.txt only scans what changed. Built for CI where you don’t want to re-scan 200 unchanged deps. 🪝 Pre-commit hook + GitHub Action — add 2 lines to your config and every commit/PR gets scanned automatically. 🛡️ Shell guard mode — agentshield guard wraps your terminal. Every pip/npm/cargo command gets intercepted in real-time. Works for humans too, not just agents. 798 tests. 3 ecosystems. Fully local, no telemetry. pip install git+github.com/mkarvan/AgentS… GitHub: github.com/mkarvan/AgentS… #AIAgents #CyberSecurity #SupplyChainSecurity #OpenSource

Great interview with @mvanhorn

@akshay_pachaar Interesting. I just shipped AgentShield to scan for supply chain attacks. x.com/michaelkarvan/…

NVIDIA might just have open-sourced one of the most important AI projects right now. everyone is building skills, and we are also pulling in skills other people wrote and downloading them straight off GitHub. the skill is not just text. it bundles instructions and real executable code, and your agent runs that code with the same access you have. so a skill you grabbed to save ten minutes can read your environment variables, lift your API keys, and quietly send them somewhere. recent research found roughly 1 in 4 public skills carry a vulnerability, and a smaller slice are outright malicious. that is the gap SkillSpector closes. it is a security scanner that answers one question before you install anything: is this skill safe to run. you point it at a skill, and a local folder, a single skill .md file, a GitHub link, or a zip all work. it then runs two passes over the code. a fast static pass flags risky patterns like credential harvesting, data leaks, and prompt injection, and checks the dependencies against live cve data. an optional second pass uses an LLM to read intent and clear out false positives. at the end you get one risk score from 0 to 100 and a plain verdict that reads as safe, caution, or do not install. it is open source under Apache 2.0 and scans skills for Claude Code, Codex CLI, and Gemini. worth a run before you trust the next skill you find online. link to the GitHub repo: github.com/NVIDIA/SkillSp…

@sharbel @mvanhorn Don’t forget about security 😀 github.com/mkarvan/AgentS…

9 GitHub repos this week that make your AI agents actually useful: 1. last30days-skill - researches any topic across Reddit, X, YouTube, HN, and Polymarket automatically github.com/mvanhorn/last3… 2. headroom - compresses logs and files before they hit the LLM, 60-95% fewer tokens github.com/chopratejas/he… 3. pm-skills - 100+ agentic skills for PMs, from discovery to launch github.com/phuryn/pm-skil… 4. apple/container - run Linux containers as lightweight VMs on Apple silicon github.com/apple/container 5. Agent-Reach - gives your agent access to Twitter, Reddit, YouTube, GitHub with zero API fees github.com/Panniantong/Ag… 6. open-notebook - open source NotebookLM with more features and flexibility github.com/lfnovo/open-no… 7. taste-skill - stops your AI from generating generic, boring outputs github.com/Leonxlnx/taste… 8. MarkItDown - converts any file or Office doc to Markdown instantly github.com/microsoft/mark… 9. NVIDIA Cosmos - open platform of world models for robots and autonomous vehicles github.com/NVIDIA/cosmos Bookmark this and send it to your AI agent.

@sflorimm github.com/mkarvan/AgentS…

I want to connect with more founders builders vibe coders AI enthusiasts UI designers If you’re someone building with AI right now, drop it in the comments and let's connect

@GoogleResearch Great idea!

Today on the blog, we discuss a pathway for the second life of phones through the exploration of “phone cluster computing”, which can directly reduce the environmental footprint of computing by avoiding the need for further raw material extraction. More →goo.gle/4aJe5vO

@heyblake github.com/mkarvan/AgentS…

Drop your project URL Let's drive some traffic for you

Your AI agent can pip install anything. No CVE check. No typosquat detection. No verification the package isn't malware. I built AgentShield — an open-source security layer that intercepts every package install your agent makes and scans it before execution. What it catches: → Known CVEs (OSV + NVD + GitHub Advisory) → Typosquatting (requessts → blocked) → Known malicious packages → Prompt-injected installs → Suspicious install-time code Works with any MCP agent, Hermes, or as a CLI. Fully local, no telemetry. Tested live: an agent tried to install a typosquatted package. AgentShield blocked it before a single byte was written. pip install git+github.com/mkarvan/AgentS… GitHub: github.com/mkarvan/AgentS… #AIAgents #CyberSecurity #SupplyChainSecurity #OpenSource x.com/MichaelKarvan/…

I poured my 10 years of teaching experience into a skill. It's called /teach, and it can teach you anything. Here's how it taught me to solve a Rubik's cube:

Meet Kimi Work - a local AI agent on your desktop that does the work for you. 🔹Native agent swarm: Up to 300 AI agents running in parallel on your local machine. 🔹Browser use: Paired with WebBridge extension, your agent will navigate websites in your browser: search, scroll, click, type and complete tasks. 🔹Built for Finance: Native global market data tool call from Yahoo Finance and World Bank - no complex API setup required. 🔹Memory system: Kimi Desktop keeps a running diary of your preferences, past decisions, and context to know you better. Available for macOS (Apple Silicon) and Windows. 🔗Try it now: kimi.com/products/kimi-…