Securing the era of AI agents
Enterprises are deploying autonomous AI agents faster than they can govern them. The data from early 2026 paints a stark picture: production is way ahead of security, identity systems built for humans are buckling under machine-scale access, and most organizations have already been burned. Here’s what the numbers say and what it means for founders building in this space.
For most of the last two years, “AI agent economy” has been a thesis about capability what agents can plan, coordinate, and execute. In 2026, a second thesis has caught up fast: what happens when you give software that much access, that fast, with identity and permission systems that were designed for humans logging in with a password. The early data on this isn’t subtle. Agentic AI is now the top-ranked security concern heading into 2026 ahead of deepfakes, ahead of board-level cyber risk, ahead of passwordless adoption.
Adoption is outpacing control by a lot
The headline number from Gravitee’s State of AI Agent Security 2026 report, based on a survey of over 900 executives and technical practitioners, is that AI agents are no longer experimental. 80.9% of technical teams have moved past the planning phase into active testing or production. Separately, Teleport’s 2026 infrastructure security report based on 205 CISOs and security architects found that 70% of enterprises already have agents running in production.
The problem is what’s not keeping pace. In that same Gravitee survey, only 14.4% of organizations report that all of their AI agents went live with full security and IT approval. Most production deployments, in other words, happened with security playing catch-up or not in the room at all.
That last figure is worth sitting with. Only 3% of organizations have automated, machine-speed controls governing how their AI agents actually behave once deployed. Everything else incident detection, permission review, anomaly response is still happening on human timescales, for systems that operate on machine timescales. As Teleport’s CEO Ev Kontsevoy put it, the issue isn’t that the AI itself is unsafe. It’s the access being granted to it.
The identity crisis underneath it all
Nearly every report on this topic in early 2026 converges on the same root cause: identity. Every AI agent introduced into an organization is a new non-human identity that needs API access and machine-to-machine authentication and most identity and access management (IAM) systems were never built with this in mind.
Read that second row again: 43% of organizations rely on shared service accounts for AI agents the security equivalent of an entire engineering team logging in with one password. And nearly a third let agents operate under a human employee’s own identity, meaning an agent’s actions and a person’s actions become indistinguishable in the logs. The Cloud Security Alliance found that 73% of organizations expect AI agents to become vital to their operations within a year, while 68% can’t clearly distinguish between human and AI agent activity in their systems today.
KPMG’s Cybersecurity Considerations 2026 report frames the scale of this starkly: non-human identities now outnumber human users by roughly 80 to 1 in the average enterprise, and in some sectors that ratio exceeds 110 to 1. Traditional identity governance built around manual onboarding and periodic access reviews simply wasn’t designed for a population of identities that size, let alone one that’s growing autonomously.
Why agents are a different kind of risk
It’s worth being precise about why this is structurally different from prior waves of enterprise software risk including the “AI risk” conversations of 2023 and 2024, which were mostly about model outputs (hallucinations, bias, IP). Agentic systems introduce a new category because they don’t just generate content they take actions: executing tasks, querying databases, moving files, calling other systems and other agents.
Elevated, often invisible permissions
Teleport’s research found that 70% of enterprises with agents in production report those agents have more access than equivalent human roles often because it’s easier to grant broad permissions once than to scope them precisely for every task an agent might need to perform.
Machine-speed action, human-speed oversight
Darktrace’s 2026 survey of 1,500+ security leaders found 14% have already given AI systems the ability to remediate incidents independently, with no human in the loop while 92% say AI-powered threats are forcing a fundamental overhaul of their defenses.
There’s also a supply-chain dimension that’s specific to how agentic systems get built. The rise of low-code and “vibe coding” platforms means a much broader set of people inside an organization not just security-trained engineers are now builders with elevated access. CyberArk’s research points to incidents like the Shai-Hulud NPM worm as early examples of threat actors targeting the trust placed in developers and the package ecosystems agents are built on. And separately, shadow AI unsanctioned tools employees bring into the workplace is projected to overtake shadow IT as the top breach risk in 2026, with more than a third of breaches already involving unmanaged shadow data.
Governance is catching up slowly
The encouraging part of the picture is that governance frameworks for agentic systems are no longer hypothetical they’re being published in real time, even if adoption lags behind. A few worth knowing if you’re building or investing in this space:
Created in 2025, structures autonomous agent deployment around four principles decision sovereignty, governance by design, ultimate human control, and traceable accountability plus 18 governance KPIs.
Published January 2026, extending Singapore’s existing AI guidelines to address agent-specific risks including delegation chains and multi-agent coordination.
Applies zero-trust principles verify explicitly, assume breach, least privilege specifically to AI agent identity and access.
A draft concept paper on how agents are authenticated, how permissions are scoped, and how actions are logged public comment period ran through April 2, 2026.
92% of tech executives in the KPMG survey believe that managing AI agents will be the defining skill for security teams over the next five years. That’s not a niche specialization it’s a statement that identity and access management for machines is becoming as central to enterprise security as it has historically been for humans.
Where this is being built: a regional view
Most of the data above comes from US and European enterprises, but the build-out of AI agent security isn’t confined to Silicon Valley. CMC Global, a member of CMC Corporation and one of Vietnam’s largest IT services groups, is a useful example of how this gap is being addressed from the delivery side. CMC Global already operates a dedicated Security Operations Center (SOC) alongside a growing Artificial Intelligence Solutions practice spanning computer vision, NLP and large language models, and data science putting AI delivery and security monitoring inside the same organization rather than as separate vendors.
That combination matters for exactly the reason this post has been making: agent security is as much an identity and operations problem as a model problem. In late May 2026, CMC Global signed an AI memorandum of understanding with Silicon Tech AI as part of a Vietnam– Thailand business forum, aimed at joint software, digital transformation, and AI application development for the Thai market. A few weeks earlier, CMC proposed expanding Vietnam–Korea cooperation specifically across AI, semiconductors, and cybersecurity three areas that, per the data above, are converging fast.
For founders and investors thinking about the AI agent security category, this is a useful data point on where supply is forming, not just where demand is loudest. North American and European enterprises are generating the survey data on agent incidents and identity gaps; delivery organizations in Vietnam, Korea, and across APAC with existing SOC infrastructure, AI practices, and cross-border partnerships already in motion are positioned to build a meaningful share of the tooling and managed services that close those gaps, particularly for the large population of mid-market enterprises that can’t build agent security capabilities in-house from scratch.
What this means for builders and investors
I’ve written before about the AI Agent Economy thesis multi-agent infrastructure, vertical domain agents, and governance frameworks as three of the most interesting areas for early-stage investment. The data above sharpens why governance, specifically, isn’t a “nice to have” layer bolted on after the fact. It’s quickly becoming a prerequisite for enterprise deployment at all.
For founders selling into the enterprise
If your product is an agent or orchestrates agents expect security and identity questions to show up earlier in sales cycles than they used to. With 88% of enterprises already reporting an agent related incident, buyers have war stories. Products that ship with clear identity scoping, audit trails, and permission boundaries by default have a real differentiator, not just a compliance checkbox.
For infrastructure and tooling founders
The gap between 70% production adoption and 3% automated behavioral controls is, from an investment lens, a massive whitespace. Identity-for-agents, agent-specific observability, and machine-speed policy enforcement all sit squarely inside categories where demand is already proven and supply is early.
For investors doing diligence
For any portfolio company deploying agents internally not just AI-native startups it’s worth asking the same questions the surveys above are asking at the macro level: are agents treated as independent identities, what’s their permission scope relative to human roles, and who reviews their actions, on what cadence?
For everyone: watch the regulators
Frameworks from Singapore’s IMDA, NIST, and industry bodies like CSA are moving from draft to adoption faster than prior AI governance cycles. Cross-border companies especially those operating between the US and Asia should expect agent-specific compliance requirements to arrive sooner than the general pace of AI regulation might suggest.
The bottom line
The AI Agent Economy thesis has always been about more capable systems doing more consequential work. 2026’s data is the other half of that story: capability and access have outrun governance, and the gap is now large enough to be the top-ranked concern among security professionals. That’s not an argument against the agent economy it’s an argument that the companies building the identity, access, and oversight layer for agents are solving one of the most urgent infrastructure problems in enterprise technology today, whether or not “AI” is in their name.
Disclosure. The author serves as a Senior Advisor to CMC Global, referenced above.