Snyk

Eleven years ago, three engineers had one idea: put security in the hands of the people writing the code. That idea is now a platform securing code, dependencies, containers, and cloud for thousands of teams worldwide. None of it happens without the people who built it. Thank you to every Snyker who shipped a feature, closed a deal, or believed in this mission before it was obvious. Eleven years of building fast and staying secure. Now we're doing it for the age of AI, securing the agents and AI-generated code, shaping how software gets built next. Join us on our next chapter! Check out open roles 👉 https://snyk.io/careers/

Anthropic Head of Applied AI Alon Krifcher and Snyk's CTO & Chief Innovation Officer Manoj Nair and sat down to discuss the coming wave of autonomous attacks. Attackers can now find architectural flaws at machine speed, and AI is widening the target at the same time. Legacy bugs, new code, and agents in production are all exposed at once. What has changed at the model layer, how it lands in production, and where Anthropic's job ends and yours begins. Plus a 10-minute Continuous Offensive Security demo. Watch below for s sneak peak and see the full session on demand here: https://vist.ly/5jfb9

Snyk is heading to the The AI Security Summit in San Francisco on October 15 🌁 AI is shipping code faster than most security teams can keep pace with, and this is the one event built entirely around closing that gap. Want to join us? Register here: https://vist.ly/5j9us

Snyk is in the right place at the right time as the market shifts rapidly to agentic AI security. Evo by Snyk has sustained 81.5% month-over-month customer growth since its first capability reached general availability in March. The number that says even more is the 76% of enterprises who bought Evo last quarter and had it running in production before the quarter closed. Enterprises aren't piloting this. They're moving as fast as the risk is. That urgency has a cause. Across our 4,800+ customers, newly introduced vulnerabilities climbed 108% between Q4 2025 and Q1 2026, and that was before agentic code generation reached full scale. Agents widen the gap at both ends, generating more code than humans can review while attackers probe at machine speed, and no amount of AI grading its own homework closes a gap like that. It takes independent validation: a deterministic layer beneath the model with more than a decade of context. We built Evo for exactly this moment. Read the full story here: https://vist.ly/5itfu

Anthropic's own threat intelligence report, published last week, describes an attacker whose malware got caught by security tools — then had it rewritten and redeployed by its own AI agents until it evaded detection again. In hours, not weeks. AI has changed the attack landscape. That's the gap we're getting into on tomorrow. The Register's James Hayes talks with Snyk's Manoj Nair and Anthropic's Alon Krifcher about closing it in four moves: discover the full attack surface, remediate the backlog, validate what's actually exploitable, prevent what comes next, seen from two vantage points on the same problem, the model layer and the application layer, Register here: https://vist.ly/5iqq5

Ask an agent to "write secure code" and it will try. But an instruction isn't a constraint. It shifts what gets written, it doesn't decide it 🔏 Read Ezra Tanzer's piece on why prevention for agent-generated code isn't one control. It's four, and each one only works at a specific point in the loop: guidance in the agent's context for the risks you can name in advance, deterministic scans that fire while the agent still holds the prompt, semantic review of the assembled codebase for the logic flaws no scanner catches, and adversarial testing against what's actually running. Read more here: https://vist.ly/5ij58

This weekend produced the most candid moment in AI safety yet. Anthropic published a threat intelligence report on how its models have been misused, and Dario Amodei called on the industry to pace the frontier via embedded evaluators, coordinated safety standards, and a deliberate slowdown in capability gains. Sam Altman agreed within hours. AI stocks fell Monday; cybersecurity stocks rallied. Pacing the frontier is a good idea, though an incomplete one. It governs what happens before a model ships. It says far less about what happens after inside the codebases and agent workflows already running these models in production, with real credentials and real infrastructure attached. Anthropic's own report describes attackers stealing AI credentials and using them as compute, cover, and a way into someone else's environment. That's a software supply chain problem as much as a frontier one, and it's already showing up. My colleague Manoj put it well: this is hurricane season, and most organizations don't even know if their shutters close. Rules built for the frontier take time to write and time to bite, and they aren't built to reach where most of the exposure lives. That takes a warning system built somewhere else, and it comes down to three things that have nothing to do with model capability: 1. The check can't come from whoever built the system. Anthropic just applied that logic to model evaluation. It applies just as much to everything running underneath the model. Independent validation cannot be overlooked. 2. Most organizations can't say what AI is actually running in their own environment — which models, which agents, which of their own code a machine wrote. That blind spot is where exposure starts. 3. Security used to run on a quarterly cadence because that's how fast software changed. It doesn't anymore, and the review process meant to catch risk can't either. At some point this stops being a debate and becomes a decision: security comes first. Not after the incident. Not after the regulation. First.

The AI hurricane is here. Debating its speed will not “weatherproof” the software we already run or protect against the attackers who aren’t slowing down. Creation moved to machine speed. Validation did not. Meanwhile, agents are writing code, choosing tools, and acting in production—and attackers can exploit that growing exposure with AI of their own. Dario & Sam’s call for independent lab evaluators and George Kurtz’s answer is the operational one: the frontier moves at whatever speed it moves, but the rest of the world does not slow down. Both are right. But we have to take it one layer upstream. I’d extend that responsibility into development. ❓Who governs the agents doing the writing? ❓Who checks the tools, skills and MCP servers they trust? ❓Who independently validates the code, and the fixes they ship? The system creating a change cannot be its own sole validator. We need independently executed tests, production evidence, and validation controls the generating agent cannot alter. And defenders need the means to act. Open models, shared intelligence, and supported open-source maintainers belong inside the safety architecture. Continuous defense must be affordable and broadly accessible. ✅ Secure at inception. ✅ Enforce at runtime. ✅ Validate independently. Read my in-depth thoughts and recommendations here, along with the questions leaders should ask their teams (or their agents) now. https://lnkd.in/eiyiuxuj The strongest movement is still Open Source and Open Models. The Security companies are Allied. The Defenders are still standing strong. Open. Allied. Still Standing - thats the mantra to focus on the most important task right now and cut through the opera.

"Shift in" is the new "shift left." But how far in can you realistically shift? As an industry we have the tools necessary for pre-deployment security. What most companies don't have yet (while development keeps moving toward more autonomous loops) is a lifecycle stable enough to instrument end to end. I've been writing about this, and where a control belongs comes down to two things at any given loop: what the agent knows and how long it has to act. https://lnkd.in/gUtrZVQm Which part of your (AI)SDLC has proved stable enough to introduce security into, and what's worked so far?

Most security programs weren't built for an attacker that moves at machine speed, and next week is when that gap gets put on the table. Next week, The Register's JAMES HAYES talks with Snyk's Manoj Nair and Anthropic's Alon Krifcher about the gap most security programs aren't built to close: AI finding architectural flaws faster than anyone can patch them, while that same AI ships code faster than anyone can review it. Closing that gap comes down to four moves: discover the full attack surface, remediate the backlog, validate what's actually exploitable, prevent what comes next, seen from two vantage points on the same problem, the model layer and the application layer. If your program still runs at ticket speed, this is the session. Register here: https://vist.ly/5hz4y