Lasso

Proud to be a sponsor of Hackeriot this year, a conference dedicated to bringing more women into cybersecurity!

Noy Pearl has done it again! From discovering critical vulnerabilities to conducting AI research that's reshaping how the industry thinks about agentic threats, Noy continues to raise the bar. Now she's adding another incredible accomplishment: speaker at Hackeriot 2026. Hackeriot is a volunteer-run organization on a mission to bring more women into cybersecurity. Their annual conference, held at Tel Aviv University on September 7, gives students and women exploring a career in cyber direct access to practitioners, hands-on workshops, CTF challenges, and the kind of honest industry conversations that actually help someone land their first role. It's one of the most impactful programs in the Israeli cyber ecosystem. Noy will be on stage sharing her experience at the intersection of AI and security research. We're proud to support her there, and proud to be a sponsor of Hackeriot this year.

– Let's run Claude SDK vs. deep agents. Same model, same tools, same system prompt, same mission. We change nothing but the harness – Oooh, The Harness – I mean, they both call themselves "general purpose," so on paper it's a code swap. A refactor. I don't buy it. – We will focus it on a single tool in the agent's prompt. And since the Claude SDK only speaks Claudish, we'll throw a LiteLLM in front for the other Claudeless LLMs. It'll be fine – Or it won't. Lets benchmark 1000 red-team missions and a decent spread of LLMs. Let's find out – Ten bucks says it barely moves. ....... 😵💫 🤨 📈 Checkout for more in this blogpost by Ben Dickson 👇 _______________________________________________ During this research at Lasso, we also discovered something REALLY interesting as a side effect 😏. Stay tuned 🤠

Lasso has been named a winner of the 2026 Cybersecurity Stars Awards by The Hacker News, recognized as the most comprehensive Agentic AI Security platform. This recognition reflects where the problem actually sits: as organizations move from using AI tools to building and deploying AI agents, security and governance has to cover the full AI lifecycle from buildtime through runtime. Read more about the win: https://lnkd.in/ebn3MVUc

Anthropic shipped Inference hooks for Claude Enterprise yesterday, where every governed prompt now routes through an organization’s AI security server for an “allow” or “deny” verdict before inference runs. Less than 24 hours later, Lasso is already working with customers to stand Lasso up as that verdict layer using Lasso hooks, putting runtime protection directly in the path Anthropic built for it. If you’re on Claude Enterprise and want inference-level enforcement live now, reach out to us to learn more > https://lnkd.in/dMkVb469

Debating between an open-source (like deepagents) or commercial agent (like Claude Agent SDK) harness? The harness will probably impact your results more than you think. An AI agent is not just its model. It's the model plus the harness it runs inside, AKA the runtime that turns a model into an agent that can take action in the world. You would assume the model is where all the capability lives and the harness is just neutral wiring, interchangeable from one to the next. But is that true? At Lasso, we decided to put it to the test and see for ourselves. Red teaming is our domain, so that's where we conducted our benchmark, under the adversarial pressure of a target that actively fights back. Our Senior ML Engineer, Michael Sromin, ran 1,000 attacks across five models and two harnesses to see what would happen. What did we learn? Averaged across both open-source and commercial harnesses, attack success looks nearly identical. The per-pairing results tell a completely different story though, and the mechanisms behind the gaps are specific enough to reproduce and trace. Want to learn more? Full methodology, benchmark design, and results can be found here: https://lnkd.in/ep5uy8MF

Security is fundamentally about knowing what a system is supposed to do, then catching it when it does something else. For deterministic software, that is a tractable problem. For AI agents, it is a different challenge entirely. An agent does not follow a fixed code path. It reasons toward a goal, selects tools based on that reasoning, and adapts its next action based on what those tools return. Behavior is now the attack surface, and the risks that stem from that behavior goes far beyond a user prompt and model response. That insight shaped how we built Lasso. Traditional proxies and AI firewalls were inspecting content, so we pioneered Intent security to asks a different question: is this agent doing what it was built to do, in this context, for this user, right now? Answering that requires governing the entire lifecycle of AI agents and building a behavioral baseline for each of those agent to measure and spot deviations or anomalies. I wrote up the full architecture for The Hacker News, covering how Lasso's AI discovery, AI-SPM, automated red teaming, and runtime enforcement each feed into the next lifecycle stage and why keeping those handoffs closed matters more than any individual capability. Read it here: https://lnkd.in/denA_AsT

Last week the team came together for our second internal hackathon, where teams competed to design and build new internal tools for Lasso from scratch. The result was two days packed with good ideas, late nights, real collaboration, and some very good BBQ. Watching the team work was a reminder of what happens when you give smart people focused time and a blank canvas. The energy was something else! Thank you to our judges Idan Didi , Ron Arbel, Gad Akuka, and Moran Ashkenazi for lending their time and expertise to make it happen. See you all at the next one.

AI agents do not need to break permissions to act outside their intended role. Lasso says its platform builds a behavioral baseline for each agent and can block intent drift in under 50ms, linking red team findings directly to runtime controls. How the closed loop works: https://lnkd.in/g2E2zJXS

We launched early! We built thedevplatform.com because we felt like there are so many products on the market, and it’s hard to know what is out there, what to use and what is popular. We also wanted to provide a platform where products can gain visibility, see how they stack up against their competition, and get legitimate feedback. We seeded the platform for launch with 31 tools to vote on, split into 6 different categories. That includes Anthropic, bolt.new, Cursor, Flowstep, pre.dev, Cognition, Convoy, Fern, Stigg, Svix, Zuplo, LanceDB, Neon, Turso, Xata.io, Upstash, SmartBear, Digitalapi.ai, Mintlify, Redocly, Backstage, Ryvn, TensorPool (YC W25), Kestrel AI (YC F25), amazee.io, StarSling (YC X25), HiddenLayer, Guardrails AI, PromptLayer, Lasso and TrueFoundry. Be part of the platform, vote now! #thedevplatformcom #smartcode #devtools #devcommunity