Liran Tal

If you're not concerned of malicious and insecure SKILL md files you should be and maybe I'll convince you in this presentation of mine form AI DevCon: https://www.youtube.com/watch?v=oJGX8GYLWxg&list=PLISstAySqk7KtlYPFps1ZnK9xNIMSVkDD&index=37 In this session you'll watch live hacking of a malicious skill and how it fools a coding agent for rogue actions, a prompt injection leaks your secrets over email, and a leaky skill passes credit card numbers straight through the LLM context. Then we flip to defense.

guess what really unlocks agentic loops? verifiable software ;-) yep, you still need to write tests, have a deterministic CI setup, and... security controls

jack of all trades, orchestrator top gun

wip for running the benchmarks on VulnBench 2.0 which is a new set of fixture data, larger apps codebase, varied language ecosystem... pretty interesting how harness + model are very much a pair, for example Codex Security agent with Terra on xhigh just isn't scoring high enough

The OpenJS Foundation contributors leaderboard 🎉 Thank you to all of the amazing humans, some of whom are on this list that I am grateful to call friends, for building a better open-source world for us ❤️

that empty real estate space on the left-side of the X feed on the web is a perfect placement for my Codex pet apparently 😆

Do you want to join my colleague Vandana who received her 𝗔𝗜 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀 certificate? Here's everything you'll learn as the baseline for AI security engineering: ✅ OWASP Top 10 for LLM Applications ✅ Addressing Shadow AI ✅ AI Threat Modeling ✅ Securing Agent Skills and MCP ✅ Securing Vibe Coding ✅ AI Red Teaming We created the curriculum, you should go take the training (on-demand)!

lol AI overview

it's 2026 and Atlassian didn't add export to Markdown in Confluence :(

working with a second brain is underrated if you're not yet there, you should commit to this now

I was really hoping to use VS Code for managing the second brain Obsidian-like vault but the extensions and overall capabilities are just terrible What did you all do? Just use Obsidian?

the pressure is real 😬

Who from the Codex team wants to confirm? 😁 Codex app doesn't want us to use emojis or em-dash. Ok the last one I can get behind. Emojis though is what I instruct coding agents to use in a CONTRIBUTING.md file so I can detect AI contributions

Black Hat 2026 will be riddled with AI Security announcements but this one from Snyk I am genuinely excited about Agentic Development Security is how we started with shaping GenAI code to ensure secure and trusted output and then continued to MCP security and then Skill scanning

hah, I do appreciate this update from dependabot but like 10 years too late Snyk dependency upgrade introduced a 21 days cooldown period since 2020 in my repos I have this defaulting to 30 days. call me paranoid. I know.

if you haven't yet jumped on the LifeOS wagon, well... there's no better time. go build yourself a second brain.

color me surprised 😅 ladies and gents, always push back

you know why, right?

let's talk about this planning skill (and in general), when they rely on pre-determined plan filenames then that mandates using worktrees, otherwise, other sessions will share the same plan files and you can't multi-task... what's your way of working for planning with agents?

appreciate the shoutout from Yigitcan Kaya on the Real World AI Security conference stage from a couple of weeks back on Snyk's ToxicSkills research malicious skills research is incredibly relevant and I'm grateful to have had the privilege to work alongside smart humans at Snyk who pioneered this work around agent skill security

bugs detected per lines of code (from the Cubic AI report) is not a bad matrix to evaluate code quality from GenAI quite interesting to see Opus 4.8 does so terribly bad sgainst GPT 5.5

we're in 2026, AI-pilled and the right-click spelling menu in macOS is horrible and useless, what the actual hell

what AX means

I got an existential question... where do you place your Codex pet on the screen? 😅

looks like a lot of you are working with agents in a highly privileged mode + highly sensitive environment (your local dev machine) how do you sleep at nights? 😅

New npq feature 🎉 To help you with figuring out an alternative npm package when the latest one doesn't fit the security criteria, npq will not check prior versions published and suggest them instead Thank you to Brian Clark who suggested this improvement npm install -g npq@latest to get it

it's been cooking for quite some time but it's not the duration but rather that it's a relatively small scope so now I'm super curious at what the agent stumbled onto throughout that required this amount of work to be invested... how do you all do agents observability locally?

you're all rushing into Skynet speculations from Mythos and other cybersecurity models but in the meanwhile, in the reality of things, about 3 years later after I disclosed these set of (CVE) vulnerabilities, I received a note that a maintainer will look into addressing them in a fork 🤷‍♂️

oh my god, new hell discovered

DevRel friends you should invest more in doing deep research I ran the VulnBench benchmark project at Snyk, which now graduated to a full-time Forward Research initiative under DevRel, and it's not only satisfying from a technical perspective but also incredibly energising to dive into the details

this is what going agentic looks like, hockey stick all the way up ;-)

ok what, petdex is really cool :)

Something with the Codex Sandbox setup (not sure if due to the new ChatGPT superapp or not) is kinda messed up

Can I trick Gemini CLI to pop the calc? apparently yes: https://lirantal.com/blog/gemini-cli-invisible-unicode-skill-injection I disclosed a Gemini CLI prompt injection where invisible Unicode tag characters in SKILL md caused the agent to run a hidden command. Google closed the report as out of scope because the workspace was trusted and YOLO mode was enabled.

Quick tip - in the latest update of Claude Cowork you can now select text on a doc (the opened right sidebar view) and then tell Claude what to change on it specifically

When Claude Code prints that token stat (10k tokens) I assume those are output tokens, right? how do you know?

from CLI to an isolated workspace environment I can work with Claude Code app in under a minute. Slick DX powered by Boxdown ✨

my codex pet is kinda cute, I like rocky, he's cool but I do think I'm gonna create my own guess what my new Codex pet is going to be lol

in case you were living under a rock and missed the OpenAI + Hugging Face incident from last week... Here, I highlighted everything you need to know from the incident report. Maybe emergent properties? Maybe Skynet early days. Maybe a nice PR. Either way, stay vigilant.

how long until we stop saying "AI agent" and just say "agent" ? I call 2 months

vanity metrics of the day

if you like this tshirt then shout out to Chris Suen who coined the trademark in a rando Slack conversation hit us up at Snyk and I'll show you what this means! ;-)

btw you can just open issues for gratitude

Boxdown vs NVIDIA OpenShell assessment for the base Docker image setup

would be nice to have the ChatGPT app support keyboard mappings for model definitions (model + reasoning + speed) so I can easily toggle between them you know ;-)

what if I ran /grill-me but I'd have the agent answer its own questions

models should be steered, then challenged, repeat. that's the real loop. real case from working with Codex (GPT 5.5 Extra High)

Snyk VulnBench JS 1.0 shows that the best-scoring LLM configuration reached 75.4% Snyk-reference F1, leaving a 24.6-point gap against deterministic SAST reference reproduction. Go read up on this coding agents security benchmark

secure software factory >> software factory

A few weeks back we announced VulnBench at Snyk. It's a forward research benchmark project that I lead to uncover LLM security findings The benchmark isn't about a competition between AI-based scanning to SAST-based scanning, but rather centers on the differences and risks of both