VentureBeat just lately sat down (just about) with Itamar Golan, co-founder and CEO of Prompt Security, to speak via the GenAI safety challenges organizations of all sizes face.

We talked about shadow AI sprawl, the strategic selections that led Golan to pursue constructing a market-leading platform versus competing on options, and a real-world incident that crystallized why defending AI purposes isn't optionally available anymore. Golan offered an unvarnished view of the corporate's mission to empower enterprises to undertake AI securely, and the way that imaginative and prescient led to SentinelOne's estimated $250 million acquisition in August 2025.

Golan's path to founding Immediate Safety started with educational work on transformer architectures, nicely earlier than they turned foundational to at present's massive language fashions. His expertise constructing one of many earliest GenAI-powered safety features utilizing GPT-2 and GPT-3 satisfied him that LLM-driven purposes have been creating a wholly new assault floor. He based Immediate Safety in August 2023, raised $23 million throughout two rounds, constructed a 50-person workforce, and achieved a profitable exit in beneath two years.

The timing of our dialog couldn’t be higher. VentureBeat evaluation exhibits shadow AI now prices enterprises $4.63 million per breach, 16% above common, but 97% of breached organizations lack primary AI entry controls, in accordance with IBM's 2025 data. VentureBeat estimates that shadow AI apps may double by mid-2026 based mostly on present 5% month-to-month development charges. Cyberhaven knowledge reveals 73.8% of ChatGPT office accounts are unauthorized, and enterprise AI utilization has grown 61x in simply 24 months. As Golan instructed VentureBeat in earlier protection, "We see 50 new AI apps a day, and we've already cataloged over 12,000. Round 40% of those default to coaching on any knowledge you feed them, which means your mental property can change into a part of their fashions."

The next has been edited for readability and size.

VentureBeat: What made you acknowledge that GenAI safety wanted a devoted firm when most enterprises have been nonetheless determining the way to deploy their first LLMs? Was there a selected second, buyer dialog, or assault sample you noticed that satisfied you this was a fundable, venture-scale alternative?

Itamar Golan: From an early age, I used to be drawn to arithmetic, knowledge, and the rising world of synthetic intelligence. That curiosity formed my educational path, culminating in a research on transformer architectures, nicely earlier than they turned foundational to at present's massive language fashions. My ardour for AI additionally guided my early profession as a knowledge scientist, the place my work more and more intersected with cybersecurity.

Every thing accelerated with the discharge of the primary OpenAI API. Round that point, as a part of my earlier job, I teamed up with Lior Drihem, who would later change into my co-founder and Immediate Safety's CTO. Collectively, we constructed one of many earliest safety features powered by generative AI, utilizing GPT-2 and GPT-3 to generate contextual, actionable remediation steps for safety alerts. This decreased the time safety groups wanted to grasp and resolve points.

That have made it clear that purposes powered by GPT-like fashions have been opening a wholly new and susceptible assault floor. Recognizing this shift, we based Immediate Safety in August 2023 to deal with these rising dangers. Our objective was to empower organizations to journey this wave of innovation and unleash the potential of AI with out it changing into a safety and governance nightmare.

Immediate Safety turned recognized for immediate injection protection, however you have been fixing a broader set of GenAI safety challenges. Stroll me via the complete scope of what the platform addressed: knowledge leakage, mannequin governance, compliance, pink teaming, no matter else. What capabilities ended up resonating most with prospects that will have shocked you?

From the start, we designed Immediate Safety to cowl a broad vary of use instances. Focusing solely on worker monitoring or prompt-injection safety for inner AI purposes was by no means sufficient. To really give safety groups the arrogance to undertake AI safely, we would have liked to guard each touchpoint throughout the group, and do all of it at runtime.

For a lot of prospects, the actual turning level was discovering simply what number of AI instruments their workers have been already utilizing. Early on, corporations usually discovered not simply ChatGPT however dozens of unmanaged AI providers in energetic use fully exterior IT's visibility. That made shadow AI discovery a crucial a part of our answer.

Equally essential was real-time sensitive-data sanitization. As an alternative of blocking AI instruments outright, we enabled workers to make use of them safely by routinely eradicating delicate data from prompts earlier than it ever reached an exterior mannequin. It struck the steadiness organizations wanted: sturdy safety with out sacrificing productiveness. Workers may hold working with AI, whereas safety groups knew that no delicate knowledge was leaking out.

What shocked many shoppers was how enabling secure utilization — fairly than proscribing it — drove sooner adoption and belief. As soon as they noticed AI as a managed, safe channel as a substitute of a forbidden one, utilization exploded responsibly.

You constructed Immediate Safety right into a market chief. What have been the 2 to 3 strategic selections that truly accelerated your development? Was it specializing in a selected vertical?

Trying again, the actual acceleration didn't come from luck or timing: It got here from a couple of deliberate decisions I made early. These decisions have been uncomfortable, costly, and slowed us down within the brief time period, however they created huge leverage over time.

First, I selected to construct a class, not a characteristic. From day one, I refused to place Immediate Safety as "simply" safety in opposition to immediate injection or knowledge leakage, as a result of I noticed that as a useless finish.

As an alternative, I framed Immediate because the AI safety management layer for the enterprise, the platform that governs how people, brokers, and purposes work together with LLMs. That call was basic, permitting us to create a finances as a substitute of combating for it, sit on the CISO desk as a strategic layer fairly than a software, and construct platform-level pricing and long-term relevance as a substitute of a slender level answer. I wasn't attempting to win a characteristic race; I used to be constructing a brand new class.

Second, I selected enterprise complexity earlier than it was snug. Whereas most startups keep away from complexity till they're pressured into it, I did the other: I constructed for enterprise deployment fashions early, together with self-hosted and hybrid; lined actual enterprise surfaces like browsers, IDEs, inner instruments, MCPs, and agentic workflows; and accepted longer cycles and extra advanced engineering in trade for credibility. It wasn't the best route, but it surely gave us one thing rivals couldn't faux: enterprise readiness earlier than the market even knew it will want it.

Third, I selected depth over logos. Slightly than chasing quantity or vainness metrics, I went deep with a smaller variety of very severe prospects, embedding ourselves into how they rolled out AI internally, how they thought of threat, coverage, and governance, and the way they deliberate long-term AI adoption. These prospects didn't simply purchase the product: they formed it. That created a product that mirrored enterprise actuality, produced proof factors that moved boardrooms and never simply safety groups, and constructed a stage of defensibility that got here from entrenchment fairly than advertising.

You have been educating the market on threats most CISOs hadn't even thought of but. How did your positioning and messaging evolve from 12 months one to the acquisition?

Within the early days, we have been educating a market that was nonetheless attempting to grasp whether or not AI adoption prolonged past a couple of workers utilizing ChatGPT for productiveness. Our positioning centered closely on consciousness, displaying CISOs that AI utilization was already sprawling throughout their organizations and that this created actual, fast dangers they hadn't accounted for.

I wasn't attempting to win a characteristic race; I used to be constructing a brand new class.

Because the market matured, our messaging shifted from "that is occurring" to "right here's the way you keep forward." CISOs now absolutely acknowledge the size of AI sprawl and know that easy URL filtering or primary controls gained't suffice. As an alternative of debating the issue, they're in search of a strategy to allow secure AI use with out the operational burden of monitoring each new software, website, copilot, or AI agent workers uncover.

By the point of the acquisition, our positioning centered on being the secure enabler: an answer that delivers visibility, safety, and governance on the velocity of AI innovation.

Our analysis exhibits that enterprises are struggling to get approvals from senior administration to deploy GenAI safety instruments. How are safety departments persuading their C-level executives to maneuver ahead?

Probably the most profitable CISOs are framing GenAI safety as a pure extension of current knowledge safety mandates, not an experimental finances line. They place it as defending the identical property, company knowledge, IP, and consumer belief, in a brand new, quickly rising channel.

What's probably the most severe GenAI safety incident or near-miss you encountered whereas constructing Immediate Safety that basically drove dwelling how crucial these protections are? How did that incident form your product roadmap or go-to-market method?

The second that crystallized all the pieces for me occurred with a big, extremely regulated firm that launched a customer-facing GenAI assist agent. This wasn't a sloppy experiment. That they had all the pieces the safety textbooks suggest: WAF, CSPM, shift-left, common pink teaming, a safe SDLC, the works. On paper, they have been doing all the pieces proper.

What they didn't absolutely account for was that the AI agent itself had change into a brand new, uncovered assault floor. Inside weeks of launch, a non-technical consumer found that by fastidiously crafting the correct dialog move (not code, not exploits, simply pure language) they might prompt-inject the agent into revealing data from different prospects' assist tickets and inner case summaries. It wasn't a nation-state attacker. It wasn't somebody with superior expertise. It was primarily a curious consumer with time and creativity. And but, via that single conversational interface, they managed to entry a few of the most delicate buyer knowledge the corporate holds.

It was each fascinating and terrifying: realizing how creativity alone may change into an exploit vector.

That was the second I actually understood what GenAI adjustments in regards to the menace mannequin. AI doesn't simply introduce new dangers, it democratizes them. It makes methods hackable by individuals who by no means had the ability set earlier than, compresses the time it takes to find exploits, and massively expands the injury radius as soon as one thing breaks. That incident validated our authentic method, and it pushed us to double down on defending AI purposes, not simply inner use. We accelerated work round:

• Runtime safety for customer-facing AI apps

• Immediate injection and context manipulation detection

• Cross-tenant knowledge leakage prevention on the mannequin interplay layer

It additionally reshaped our go-to-market. As an alternative of solely speaking about inner AI governance, we started displaying safety leaders how GenAI turns their customer-facing surfaces into high-risk, high-exposure property in a single day.

What's your position and focus now that you simply're a part of SentinelOne? How has working inside a bigger platform firm modified what you're capable of construct in comparison with working an unbiased startup? What received simpler, and what received tougher?

The main target now could be on extending AI safety throughout all the platform, bringing runtime GenAI safety, visibility, and coverage enforcement into the identical ecosystem that already secures endpoints, identities, and cloud workloads. The mission hasn't modified; the attain has.

In the end, we're constructing towards a future the place AI itself turns into a part of the protection material: not simply one thing to safe, however one thing that secures you.

The larger image

M&A exercise continues to speed up for GenAI startups which have confirmed they’ll scale to enterprise-level safety with out sacrificing accuracy or velocity. Palo Alto Networks paid $700 million for Protect AI. Tenable acquired Apex for $100 million. Cisco purchased Strong Intelligence for a reported $500 million. As Golan famous, the businesses that survive the subsequent wave of AI-enabled assaults will likely be those who embedded safety into their AI adoption technique from the start.

Submit-acquisition, Immediate Safety's capabilities will lengthen throughout SentinelOne's Singularity Platform, together with MCP gateway security between AI purposes and greater than 13,000 recognized MCP servers. Immediate Safety can be delivering model-agnostic protection throughout all main LLM suppliers, together with OpenAI, Anthropic, and Google, in addition to self-hosted or on-prem fashions as a part of the corporate's integration into the Singularity Platform.

Source link