Andre Soto

Existing IAM concepts for agent identity and auth? Are you serious? Autonomous agents work with probabilistic intent and outcomes, shift roles rapidly, cross org boundaries, may exist ephemerally for a single task, but operate at machine speeds and scale. Your existing identity infrastructure was built for humans or machines, not agents that spawn, collaborate, and disappear in seconds. We built an Agent Identity Service powered by AGNTCY | A Linux Foundation Project that delivers: 🫆Cryptographic badge generation for MCP Servers and AI Agents using A2A / MCP / OASF specs 🫆Human-in-the-loop policy authorization for critical actions with real-time mobile approvals 🫆Native integrations with Duo Security, Okta, Ory Corp, plus AGNTCY's DID-based identity provider Tamper-proof badges that prove identity and capabilities. Plugs into whatever identity infrastructure you already run. Free SaaS built on a Linux Foundation project because agent identity standards cannot be vendor-controlled. Code on GitHub. We created a currency exchange demo showing the complete workflow. Financial assistant authenticates with exchange agent, both verify cryptographic badges, human approval required for transfers. Every action traceable to specific agent identity. No more wondering which agent transferred money or called that API. No more forensic investigations when something breaks. Read the blog by Julu Panat & Hervé Muyal for details https://lnkd.in/g_MnVuQ7 Try it. Tell me what works, what doesn't.

148

Your incident response is a perfectly executed drill for a war that's already over. We've built sophisticated processes around response: runbooks, escalations, SOC workflows. Everyone knows their role and protocols activate seamlessly. The only problem? The missile was launched 30 minutes ago. While security teams perfect response playbooks, AI-powered attackers operate upstream at machine speed. Infrastructure spins up in minutes, targeting is automated, and by the time your first alert fires, the breach is operational. This reveals a fundamental mismatch. We've designed organizations to optimize for control and compliance, but the threat environment runs on speed and operates before detection even activates. Response capabilities aren't worthless, but response-first thinking assumes you'll have time. Against AI-driven attacks that deploy faster than human processes can detect, that assumption is broken. Security teams can't keep checking boxes and waiting for alerts. They need to operate upstream, detecting pre-attack staging rather than post-breach artifacts. The goal isn't replacing what you've built but layering proactive detection on top of reactive response. How is your security adapting to attack speeds that outpace human response times?

157

9 Comments

Vibe coding is already accelerating developer workflows in ways we couldn’t have imagined a few years ago. But speed without security can quickly become a liability. Hardcoded secrets, data leaks, supply chain blind spots, and entirely new classes of vulnerabilities are already appearing. The question isn’t if you’ll face these risks, it’s when -- and how ready you are to respond. The good news: with the right awareness and preparation, teams can move fast and stay secure. If you’re building with AI today or gearing up for what’s next, this is worth your time. 👉 https://lnkd.in/eaexXUSr

84

2 Comments

"We're just really good at innovating." —Matthew Prince, Cloudflare co-founder and CEO A skeptic might call a statement like this hubris...but objectively, I think it's a pretty accurate assessment of where Cloudflare is at right now. I enjoy covering Cloudflare because every earnings report or event they do feels like a glimpse into the future of the internet (and how security will play a role in that future). Their Q2'25 earnings report was super interesting. This one goes way beyond the financial numbers. Here's what stood out to me: → The "Act 1" and "Act 2" security products are delivering in interesting ways Cloudflare is world class at layering (a16z concept of starting with PLG and adding on other products and GTM channels). A couple anecdotes they gave about their security products show exactly why. The Act 1 products (WAF, DDoS, DNS) are an entry point for AI companies. They have fairly unique needs given the frequency of requests and incentives for fraud (turning stolen credit cards into tokens). 80% of them use Cloudflare. Their Act 2 (SASE/ZTNA) products made it into the Gartner visionary quadrant in the latest report. Skeptics (and competitors) would say it's still not full featured yet — but don't overlook the compounding rate of improvement and competitive win-backs. Both sets of products are showing evidence of being an initial landing point for new customers...which opens the door for both the Act 3 products and further expansion across the rest of their platform. → More AI inference hosting = more AI security upsell I've talked a bit about Cloudflare's AI inference hosting before, and it's continuing to gain momentum. The implication for their cybersecurity business is that more inference hosting sets them up for more conversions + upsell to the AI security products they offer (and more they could offer in the future). It's a fairly simple upsell – you're already hosting these AI models here, so just flip a switch to add X/Y/Z AI security features at $X/month. The nuance (and potential differentiator) here versus everyone else who wants to secure AI models is that Cloudflare is *hosting* them. TBD how much that matters, but it can't hurt. → Content protection (Act 4): security as a business enabler Cloudflare's emerging content protection business (brokering content and transactions between publishers and LLM provider traffic) could be an amazing example of security being an enabler. Fundamentally, they're using both identity and access control in creative ways to build a new business model for the web. I know this is a stretch on the traditional interpretation of "security" (and as a business model in general)...but these are the mind-bending thought experiments they give us. --- Never a dull moment with Cloudflare! Great report – lots to be excited about for the rest of the year and beyond.

70

21 Comments

While you might be patching the 47th CVE this month, NanoVM Unikernels are booting up to 2 orders of magnitude faster than Docker and running applications up to 300% faster on AWS, 200% on GCP with a far smaller attack surface. I recently sat down with Kaivalya Apte and chatted all things unikernels. Link in comments.

69

4 Comments

Many are familiar with AITM phishing, but few have seen MFA-downgrade attacks that bypass phishing-resistant methods like Passkeys, YubiKeys, or Okta FastPass. This video shows me using Evilginx with a custom phishlet to bypass a passkey via an MFA downgrade attack on Microsoft. Credit to Yehuda Smirnov for demonstrating the original approach at BlackHat. Microsoft patched his original vector, I used a slightly different approach. This isn’t a flaw in passkeys or Microsoft, but a configuration issue that affects any system allowing multiple authentication methods. If phishable MFA methods are still accepted, you’re still vulnerable — even with phishing-resistant options enabled. In large orgs, this is common due to the overhead of managing lost tokens and alternative devices. As passkeys grow more common, MFA-downgrade attacks will become the norm. Phishing kits like Tycoon already support them. Full article - https://lnkd.in/gv3zSVSm

126

3 Comments

When we started Tailscale, our target audience was… well, people like me. Developers. The kind of folks who think JSON is not only fine, but actually a pleasant way to describe network policy. It was robust, explicit, feature-rich, and, most importantly, already in our toolbelt. For years, JSON ACLs worked great for our users because most of them were engineers. But as Tailscale grew, something happened: our audience broadened. Now we have huge numbers of people in IT and security roles using Tailscale. Many IT managers are perfectly happy editing JSON by hand—but if you’ve spent your career in other parts of IT, or you’re just getting started, “here, edit this JSON file” can feel like being told your new car only has manual transmission and you’ll have to machine the gearbox yourself. That’s why the new visual editor beta is such a big deal. We didn’t rip out JSON (don’t panic, fellow code monkeys—your curly braces are safe). Instead, we’ve built a GUI that sits alongside it. You can use either one, switch between them, and nothing gets lost in translation. The same policy can be managed visually or as code, instantly. This isn’t just for businesses. The other day, our comms lead, Will, told me he loves Tailscale personally (though I suppose he’s contractually obligated to say that) but he’s avoided writing ACL policies because JSON “looked intimidating.” Now? He’s happily defining homelab policies in the visual editor, just like any enterprise IT team would. As Tailscale grows, so does the variety of people using it. The best security and configurability should be accessible to everyone, not just those fluent in a particular syntax. Usually when you add a GUI to something, you lose power and flexibility—or you’re forced to pick one mode or the other. But with Tailscale’s visual editor, you don’t have to choose. You just get the best of both worlds. There’s more coming soon to make Tailscale easier and more delightful, but this one feels particularly special: because it means the way you work is the “right” way, no matter how you work. You can read more at our announcement blog post https://lnkd.in/gd3Hbp4P If you've tried the visual editor - would love to know what you think in the comments.

99

2 Comments

👋🏻 Hope you're having a great week! What if red teams weren't just human-led—but AI-coordinated? Agent-to-Agent (A2A) communication is the next frontier in AI-driven security. We're now seeing autonomous agents collaborate like real red teamers, sharing telemetry, context, and intent to act together—in real time. Imagine this 👇🏻 🔍 Agent 1 detects a stealthy process injection 🛣 Agent 2 maps the lateral movement path 📤 Agent 3 flags potential data exfiltration 🤝 All correlate signals instantly and act as one unit This isn't just faster security—it’s coordinated decision-making at machine speed. Think of it like self-driving cars, but for security operations. But to truly make this work, agents must: 1️⃣ Communicate using low-latency, deterministic protocols (think gRPC) 2️⃣ Access shared context to eliminate blind spots 3️⃣ Operate within strict trust boundaries to avoid cascading failures At Strike, we’re engineering this into our AI-led offensive security stack—enabling autonomous triage loops and multi-agent red teaming across complex attack surfaces. ⚠️ The potential is massive—but power needs control. 👉🏻 Where should we draw the line between autonomy and oversight in cybersecurity? Have a great and secure week ahead! #AI #Cybersecurity #RedTeam #A2A #SecurityAutomation #OffensiveSecurity #Strike

85

6 Comments

We’ve sat through hundreds of security questionnaires. And we’ll say it plainly: almost none of them make sense for AI. They were written for SaaS. Software that shipped once and stayed largely unchanged. Predictable systems built on fixed deployments. Very different from AI where models update frequently, depend on external providers, and sometimes hallucinate. When we sat on the startup side, we saw how painful these reviews could be. When we sat on the enterprise side, we saw how little clarity the questions actually gave. — So we partnered with Bland, a leading voice AI platform, to publish a framework for what questionnaires should look like in the age of AI. Instead of recycling SaaS-era checkboxes, the right questions focus on what actually defines whether an AI company is safe, reliable, and ready for enterprise scale: → Data flows: tracking the full chain of how information is transmitted and stored across an AI system. → Model governance: the ability to monitor, adjust, and roll back models when behavior shifts. → Hallucinations: understanding how hallucinations are measured and mitigated. → Intellectual property: figuring out what’s defensible when foundation models are shared across the market. These are the questions that reveal real risk. And they’re absent from most enterprise reviews today. — At Delve, we’re the AI-native compliance platform automating SOC 2, ISO, HIPAA, and more for 500+ of the fastest-growing AI companies like Lovable, Cluely, and Wispr Flow. And a couple months ago, we became compliant with ISO 42001 under the newest AI regulation. AI has changed how software is built. It should also change how we evaluate it. As the company sitting between founders, enterprises, and regulators, we believe this is the standard for what comes next. Link to our full manifesto in the comments.

193

32 Comments