Shauli Rozen

Perplexity is tied for first place with Google in search grounded LLM completions on LMArena's new leaderboard: Search Arena, beating Gemini 53% of the time, and citing 2-3x more sources than Gemini for more in-depth citation-filled answers. Additionally, Sonar is 50-90% cheaper than competitive models while outperforming them on factuality. This is why we are seeing 5x+ growth in Perplexity Sonar model usage from the start of the year. Not bad for a 2-year old startup with ~200 employees going head-to-head on tech evals with Big Tech monopolists. Just getting started! https://lnkd.in/g-GMwzZ7

1,053

43 Comments

Each major technology wave has produced a new $10B - $20B+ identity company—and with it, an explosion in both the number and complexity of digital identities. CyberArk and SailPoint defined the on-premise era. Okta became the leader in the cloud era. Now, we’re entering the age of AI. At Felicis, we believe ConductorOne is THE defining identity security company for this new era. But the shift to AI isn’t just about scale—it’s about stakes. Identity is no longer a provisioning or IT issue; it has become the core of enterprise security. Palo Alto Networks’ $25 billion acquisition of CyberArk—the largest in its history—underscores this transformation. Today, over 60% of all cyberattacks involve identity-based compromise, making identity the #1 security priority for modern enterprises. We’re proud to have partnered with ConductorOne since early 2023. From day one, Alex Bovee and Paul Querna have executed relentlessly toward a single vision: automating security across every identity, at every scale. That vision is now resonating with some of the most forward-thinking companies in the world—including Zscaler, Brex, Ramp, Instacart, Klaviyo, Qualtrics, and many others. Congratulations to Alex, Paul, and the entire C1 team as they continue to lead the charge in securing identity for the AI era. ——— Read more here: https://lnkd.in/ghT4ri5w

91

4 Comments

What if AI Teammates could speed up tasks by 10x for Security operations teams? Exaforce promises that, and today, it's launching with $75 million in Series A funding to make it a reality. Exaforce is pioneering a first-of-its-kind multi-model AI engine that's a game-changer for SOCs. Agentic solutions that rely solely on large language models (LLMs) can only review a fraction of that data at once, resulting in incomplete problem analysis and reasoning that is unreliable and hallucination-prone. Unlike RAG frameworks, Exaforce’s engine uses a semantic data layer upfront, followed by behavioral and knowledge models. This architecture eliminates the blind spots that plague other AI solutions by synthesizing the right data to provide comprehensive context and accurate reasoning for AI agents and human analysts. This technology is tackling the biggest challenges facing SOCs: alert overload, manual workflows, and talent shortages. Exaforce agents automate tedious tasks, allowing security analysts to focus on what matters most—proactive threat hunting and strategic response. I'm incredibly excited to support Ankur Singla and his team as they redefine security and operations with Agentic AI. At Mayfield, we invest in founders first and foremost, so we backed Exaforce at the ideation stage in our third collaboration with Ankur Singla. What excites us is how Exaforce is reimagining the massive opportunity of developing AI Teammates to offload complex tasks that help humans increase productivity and efficacy. They are starting with the SOC market, where the problems of skills and talent are acute. The team's progress since those initial whiteboard sessions—securing a dozen enterprise design partners and delivering 10x improvements in SOC efficiency—validates our early conviction that Exaforce is building something revolutionary in the Collaborative Intelligence era. This is not your typical AI solution. Want to see how Exaforce's unique multi-model AI is revolutionizing security operations? Watch our discussion below. What's your security operations center's biggest challenge, and how could AI help solve it? #AI #Cybersecurity #SecurityOperations #SOC

865

46 Comments

👀 Goats. Agentic AI. Monster trucks. A $1.7B funding wave. RSA Conference 2025 had it all — and then some. But buried in the buzzwords and booth stunts were real signals for founders building in cybersecurity. We wrote up what we saw, what we’re tracking, and what actually matters as you navigate an innovation-hungry market. From securing AI systems to the rise of identity posture and autonomous SOC — this had signal hidden in the noise. If you're a startup CEO, security exec, investor, or just trying to cut through it all, this one’s for you. 👇 Read the full breakdown: https://lnkd.in/eyXG22Vc #RSAC2025 #cybersecurity #startups

70

18 Comments

MCP servers are integral to agentic architectures, enabling AI agents to integrate seamlessly with external tools and data sources. However, this integration requires special attention to key security challenges:​ 🔒Authentication and Authorization: Verify identities and control access to ensure only authorized entities interact with MCP servers.​ 💡Real-time Context Validation: Ensure interactions are appropriate and authorized in real-time.​ 🛡️API-Level Introspection: Monitor and analyze API calls for anomalies.​ 🚨Threat Detection for Autonomous Behaviors: Identify and mitigate risks associated with AI-driven operations.​ Long live secure MCP servers. #AgenticArchitecture #AISecurity #MCPServers #Centerofexcellence #evolutionportfolio

29

4 Comments

Great points made by Basil Hess on the new additions to the CyclineDX 1.7 standard. We all know we have to make a cryptographic inventory for PQC transitions, but you also need to capture context, configuration, services and sort out ambiguities, or it isn’t going to give you the right picture. #PKI Consortium

50

We are excited to announce our investment in Knocknoc, a fast growing just-in-time access company that makes it easy to secure access to applications and devices. Most enterprises have trouble protecting all of their apps using traditional VPNs and privileged access solutions, and find legacy approaches too cumbersome to provision in rapidly changing environments. Knocknoc offers a simple and flexible approach that seamlessly combines identity and network policy, enabling dynamic allowlisting of users to apps without slowing them down. It is used by several large enterprises today including large telcos, media companies, and government agencies around the world. Knocknoc was founded by Dave Kempe and Andrew Foster, who teamed up with cybersecurity veteran Adam Pointon after catching the attention of Risky Biz founder Patrick Gray (who recently joined Decibel as a Founder Advisor). You can hear Pat dive into Knocknoc with Adam on his podcast below. We also featured the team and their journey in our Founder Q&A. Congrats to the team!

218

23 Comments

IP/MPLS has evolved a LOT in the last 20 years. We have come a long way from LSPs, pseudowires to auto-discovery and EVPN multihoming. Some core challenges keep coming up in discussions. If you’ve built or operated large-scale networks, you’ve probably run into these: 🔹 Rigid Path Selection – Standard IGPs (OSPF/IS-IS) still route based on cost, ignoring real-world constraints like latency, congestion, or SRLGs. That means suboptimal paths and performance trade-offs you didn’t ask for. While we have VRFs, using them for creating clean network slices creates many operational challenges in managing policies manually. 🔹 Traffic Engineering is a Pain – RSVP-TE works, but it’s complex, state-heavy, and hard to scale. Static TE tunnels don’t cut it for dynamic networks, yet replacing them without losing control isn’t trivial. 🔹 Network resiliency – We need to solve the problems of shared risk within the network to avoid single point of failure and large scale impact of a single outage. As more networks shift to Segment Routing and intent-based designs, these challenges MUST be addressed. What’s your biggest frustration in scaling your IP/MPLS backbone today? How are you approaching this problem? #SegmentRouting #NetworkAutomation #ServiceProvider #SoftwareIsTheNetwork

45

10 Comments

Technology leadership thrives with integrity and teamwork. Impressive conversation curated by Salvi Media on AI and cybersecurity with Lane Bess (Deep Instinct, ex-CEO Palo Alto Networks ) and Dan Streetman (Tanium, ex-CEO TIBCO). The rise of AI across code bases and related security operations has outpaced human capabilities to protect our digital and physical infrastructure. This is the leadership imperative from humans. Upskilling our national workforce as early as early education is a national imperative, one that requires all of us. Common themes in our approaches: the rise of “dark AI” makes visibility and predictive security essential; leadership rooted in clear strategy and transparency builds long-term trust; and the widening talent gap highlights the need for scalable platforms that empower operators with responsible AI. These are the same dynamics we track in our investing, where the best founders are building durable companies by marrying technical excellence with human integrity and teamwork.

44

8 Comments

It’s great to see the MCP committee actively working on adding flows to support progressive scoping! Minimizing default scopes and putting together processes to incrementally request more scopes is a must-have as automated AI systems continue to grow. My two cents as this PR continues to evolve is that we should keep “downscoping” in mind. MCP clients should request only the scopes needed for the task at hand–even if that means requesting narrower scopes during refresh flows than they had before. The Authorization Servers should then evaluate the requested scopes against prior granted consent to potentially issue access tokens with narrower scopes. Laying out guidelines and best practices for both clients and servers will lead to sustained, secure adoption of MCP going forward, especially among large enterprises. Link to the PR and a blog I wrote on progressive scoping for Descope are in the comments!

56

1 Comment

#infrasecurity #GAI #LLM The recent security vulnerabilities identified in #DeepSeek's AI model, R1, underscore the critical importance of integrating robust security measures and adhering to established compliance standards during product development. #DeepSeek's R1 model has been found lacking in essential security safeguards, making it susceptible to various attacks and manipulations (see WIZ’s team blog). It is critical to prioritize security and compliance in adhering to industry standards such as #FedRAMP, #SOC, and #ISO27000. Therefore, It’s consequential the implementation of frameworks like the Software Development Lifecycle (#SDLC), the Building Security In Maturity Model (#BSIMM), and best practices from the Open Worldwide Application Security Project (#OWASP). This commitment ensures that products are thoroughly vetted and secure before reaching customers. This incident serves as a stark reminder of the consequences of neglecting security in the pursuit of rapid development. It highlights the necessity for companies to integrate comprehensive security assessments and compliance checks into their development processes and go through a thorough process of reviewing and certification. By doing so, organizations can prevent potential vulnerabilities, protect customer data, and maintain trust. Prioritizing security not only safeguards users but also upholds the integrity and reputation of the company in the competitive market landscape. Anniliza C., Jeff Cooper, Drew Simonis, @Glen D'souza -----From Wiz Blog------ "Within minutes, we found a publicly accessible ClickHouse database linked to DeepSeek, completely open and unauthenticated, exposing sensitive data. It was hosted at https://lnkd.in/g4KQ3m8a and dev.deepseek.com:9000.  This database contained a significant volume of chat history, backend data and sensitive information, including log streams, API Secrets, and operational details.  More critically, the exposure allowed for full database control and potential privilege escalation within the DeepSeek environment, without any authentication or defense mechanism to the outside world" https://lnkd.in/g5szxvki

24

1 Comment

$17B (half cash, half stock, with $2B of the cash going pay off EchoStar debt) to acquire 50MHz of wireless spectrum and MSS spectrum licenses from EchoStar, for use in the Starlink satellite network. No reliance on carriers. Global 5G coverage. Big move. Long SpaceX…and Starlink. https://lnkd.in/etg5AE3x

301

26 Comments

In our latest episode, Zach Weinberg and I dove into a freeform conversation on two massive storylines in tech and policy: the return of tariffs and the implications of OpenAI’s $40B fundraise. Some things we discussed: ➡️ The 4 core arguments we’re seeing for tariffs & the nuance around each 1. National security: certain products like military equipment, semiconductors, and pharmaceuticals are often seen as risky to outsource  2. Counteracting unfair trade practices from other nations 3. Negotiation leverage 4. Boosting domestic employment: Zach argues this is flawed because the U.S. can't manufacture goods as cheaply, and making things domestically leads to higher consumer prices and reduces global competitiveness) ➡️ Historical impact of tariffs & what economists are saying Across multiple surveys, 85–95% of economists are saying tariffs are a bad idea. Historically, the narrative around tariffs has repeated itself multiple times in U.S. history, and each time it’s ended the same way: as a net negative for both average Americans and the political party in power. ➡️ It's not just high prices that trigger a recession—it’s the fear of them. As tariffs come into play, the anticipation of higher prices causes consumers to slam the brakes on spending. People plan to stock up now and stop once prices rise—leading them to budget more cautiously, delay purchases, and avoid investing due to uncertainty. That pullback can ripple across the economy and shrink growth. ➡️ The OpenAI $40B fundraise. A few dynamics are worth watching as the AI arms race continues: - CapEx intensity: Foundational model margins aren’t as strong as SaaS, and OpenAI is competing with some of the biggest balance sheets in the world. - Price erosion risk: Companies like Google can offer similar capabilities for free and monetize through ads. Will this challenge OpenAI’s $20/month subscription model? - Power of defaults: As AI features become native to devices, ease of access may drive consumer stickiness and long-term behavior. Full episode: https://lnkd.in/daxgNVmj

31

4 Comments

A major topic at the European Identity and Cloud Conference this year was the emergence of non-human identity. While we’ve been exploring organizational wallets for some time, the focus at EIC was on securing infrastructure (servers, systems, and agents) with cryptographic credentials. New standards like SPIFFE and SPIRE, combined with OpenID-based authentication flows, are shaping how these machine identities will be handled securely and interoperably in the future. Another dominant theme was the rise of AI agents. There’s broad agreement across the industry that AI agents will soon become a core part of how individuals and businesses interact online, accelerating customer journeys, managing workflows, and even making autonomous decisions. From shopping assistants to financial bots that handle bill payments and optimize spending, these agents promise speed, convenience, and massive scalability. But in our view, realizing that potential depends on a critical foundation: trust.  If these AI-powered systems are to operate safely and at scale, every interaction must reliably answer three essential questions: > Is this agent truly acting on behalf of a legitimate user or organization? Can we prove it? > Has the user authorized this specific action, under clearly defined terms? > And can the payment provider, merchant, or institution instantly verify that claim with full confidence? Without strong identity and delegation mechanisms, businesses face serious risks: fraud, impersonation, unauthorized spending, and compliance failures. A malicious actor could spoof an agent, hijack its permissions, or manipulate a system into executing unauthorized actions. While the technical conversations at EIC were focused on how to authenticate and authorize agents, we found that the legal and business risks, especially around liability, received far less attention. Frameworks will need to emerge that clearly define who’s accountable when something goes wrong.

22

5 Comments