On 28 October 2026, Lightspeed Venture Partners and Amazon Web Services, in partnership with Fortune, published the fourth annual Fortune Cyber 60 list. The list recognises the most critical venture-backed cybersecurity companies in the world. This year, one category dominated the conversation: agent security and governance.
Accompanying the list was a CISO survey conducted by Wakefield Research among 200 chief information security officers at companies with at least $500 million in annual revenue. The headline finding is sobering: 77% of surveyed large enterprises have AI agents operating in production, but only 20% of those organisations maintain a comprehensive inventory of their active agents.
Put another way: four in five large companies deploying autonomous AI agents cannot say with confidence how many agents they have, what those agents are doing, or where they have access. In the physical world, this would be equivalent to operating a fleet of autonomous vehicles without knowing how many are on the road, what cargo they carry, or who holds the keys. In the digital world, the blast radius is potentially larger.
What the survey found
The Lightspeed Cyber 60 CISO Survey paints a picture of enterprises that have moved fast on agent deployment and are only now confronting the governance implications.
Agent identity and access governance is the top priority. CISOs ranked agent identity and access governance as the single most important category they plan to evaluate in the next 12 months. This is a significant shift. Until recently, agent governance was treated as a secondary concern to agent capability. The survey suggests the pendulum has swung.
CISO confidence has dropped sharply. Enterprise CISO confidence in mitigating AI-specific threats fell from 71% last year to 55% this year. The drop is not because threats have become harder to understand. It is because the attack surface has expanded dramatically in ways that existing security architecture was not designed to handle.
Agents are deeply embedded. The survey confirms that agents are no longer confined to experimental sandboxes. They are operating in cloud environments, SaaS applications and endpoint devices, executing complex workflows that span multiple databases and production systems. The notion of “testing” agents before production deployment is becoming obsolete because many agents are already in production by the time security teams become aware of them.
Why the inventory gap matters
The 77% versus 20% split is not a statistical curiosity. It is an operational vulnerability with three immediate consequences.
Compromised agents become insider threats. When an agent is hijacked or compromised, the attacker gains access to everything the agent can reach. Because agents operate continuously, chain actions together and delegate to sub-agents, a single compromised agent can move laterally through systems in ways that would be difficult for a human attacker to replicate. CISOs now rank compromised agents alongside deepfakes and AI-generated phishing as the top AI-specific threats they face.
Permitted actions combine into dangerous behaviour. Even agents that are correctly authorised can produce unexpected outcomes when their actions are chained. An agent authorised to read customer data and another authorised to send emails can, in combination, exfiltrate data without either agent individually exceeding its permissions. Traditional identity and access management systems, built for static human permissions, do not detect these compound risks.
Internal data oversharing is now a greater threat than external malware. The survey and accompanying analysis suggest that the most common agent-related incident is not a breach by an external attacker but accidental data exposure by an authorised agent. Agents ingest data at scale, and without proper controls they can unintentionally expose confidential intellectual property across systems that should remain separate.
The architecture problem
Traditional identity and access management systems were designed for human workers. They rely on interactive logins, multi-factor authentication and static permission sets. Autonomous agents break every one of these assumptions.
Agents operate continuously. They do not log in and out. They chain actions together without human intervention at each step. They delegate tasks to sub-agents, creating permission chains that are difficult to trace. They consume data at volumes that overwhelm manual audit processes.
This architectural mismatch has created demand for what industry analysts are calling the agent control plane: a new layer of enterprise security that treats agents as first-class identities, monitors their behaviour in real time, and intervenes when permitted actions combine into dangerous outcomes. Noma, which was named to the Fortune Cyber 60 for its work in this space, describes the requirement as giving security teams one control plane to stop rogue agents without slowing the business down.
What this means for enterprise buyers
For procurement and security teams evaluating agentic AI platforms, the survey creates a clear checklist.
Demand agent inventory as a baseline. Any agent platform should provide a real-time inventory of all active agents, their permissions, their recent actions and their current status. If a vendor cannot show this, they are not solving the governance problem; they are contributing to it.
Ask about runtime behaviour monitoring. Static permissions are insufficient. Buyers should ask how the platform monitors actual agent behaviour over time, detects when authorised actions combine into risky patterns, and intervenes automatically. The shift from static defence to behavioural governance is the defining architectural change in agent security.
Require an AI Bill of Materials. Just as software supply chains now require a bill of materials, agent deployments should include an inventory of models, tools, data sources and permission chains. Buyers should ask vendors whether they provide this and how it is kept current.
Plan for a hybrid workforce security model. Enterprises are moving to a world in which humans and agents share the same systems. Access control, audit trails and incident response processes all need to accommodate both. This is a bigger change than adding a new software module. It is a fundamental redesign of how enterprise security operates.
What this means for suppliers
For companies building or selling agentic AI solutions, the message is equally direct. Governance is no longer a bolt-on feature. It is a core requirement.
Enterprise buyers will increasingly evaluate agents on visibility and control, not just capability. Vendors that can demonstrate real-time inventory, behavioural monitoring and automated intervention will have a significant advantage. Vendors that treat governance as an afterthought will find procurement cycles lengthening and security reviews becoming blockers.
The funding landscape reflects this shift. The cybersecurity market has pivoted aggressively toward agentic AI defence, with $1.68 billion invested in AI-native security startups in the first eight months of 2026. The venture capital community has recognised that where there is critical infrastructure deployed without guardrails, there is a market for guardrails.
The Agentic Expo angle
Agentic Expo, at Olympia London on 23 to 24 March 2027, is where buyers and suppliers will meet to work through these questions in person. By then, the 77% production figure will almost certainly have grown. The 20% inventory figure, one hopes, will have grown faster.
For buyers attending the expo, the practical test is straightforward: can the vendor demonstrate a live inventory of agents, show behavioural monitoring in action, and explain how their platform handles the compound risk of chained agent actions? These are not theoretical questions. They are the questions that CISOs are already asking, and the answers will determine which agent platforms get budget approval in 2027.
For suppliers, the opportunity is in the gap itself. The enterprises that are deploying agents without visibility are not reckless. They are simply moving faster than their security architecture allows. The vendors that can close that gap, without forcing enterprises to slow down, will define the next phase of the agentic AI market.
Sources: Lightspeed Venture Partners and Fortune, "Cyber60 2026-2027", 28 October 2026; BriefGlance, "The AI Blindspot: Why 77% of Corporate AI Agents Are Running Rogue", 28 October 2026; PR Newswire, "Noma Named to Fourth Annual Fortune Cyber 60 List", 28 October 2026.