Overcoming IT and Security Team Resistance to AI Agent Rollouts
Security teams aren't resisting AI agents—they're blocking unverifiable risks.

Security and IT teams are blocking AI agent rollouts because the governance gap underneath those rollouts is real and documented, and that gap, not any reluctance from security, is what produces the blocks. The pace of adoption has outrun the pace of oversight, and the numbers show how wide that gap has become.
Why security teams are blocking AI agent rollouts
Gravitee's State of AI Agent Security 2026 report found that most technical teams have already moved past planning, and are now testing or in production. But only a small fraction of those agents went live with full security and IT approval. So security teams get bypassed instead of consulted, and agents reach production without sign-off from the people who are supposed to defend the organization. Cisco's State of AI Security 2026 report backs this up from a different angle: only roughly a third of organizations feel prepared to secure agentic AI applications, which puts the remaining two-thirds in the position of running AI agents they cannot properly monitor. Put those two findings together: most organizations are deploying agents faster than they can secure them, and the teams asked to approve those deployments are being asked to sign off on systems they can't see into. Calling that resistance "slow adoption" misreads what's happening. Security teams blocking these rollouts are making the only defensible call available to them, given what they can and cannot verify, and that's what makes the resistance rational rather than reflexive.
The MCP attack surface
The Model Context Protocol sits at the center of nearly every major AI agent security incident recorded in 2026, and the risks it introduces don't look like anything existing enterprise security tooling was built to catch. MCP gives agents standardized access to databases, APIs, file systems, and internal tools through a single protocol interface, and the same standardization that makes MCP powerful is what turns it into a single, scalable attack surface. These aren't hypothetical risks modeled in a lab. Check Point Research disclosed remote code execution in Claude Code through poisoned repository configuration files. Antiy CERT confirmed a large volume of malicious skills spreading across ClawHub, the OpenClaw AI agent marketplace. Trend Micro found 492 MCP servers exposed directly to the internet with no client authentication and no traffic encryption at all.
The Supabase incident shows just how little an attacker needs, as General Analysis documented it. An agent had service-role database access, and when it processed a routine user support ticket, embedded instructions told it to query an integration_tokens table and post the contents publicly. That was a proof-of-concept demonstration rather than a confirmed breach, but the mechanism it exposed is the one security teams are now worried about everywhere: no credential theft, no network intrusion, just a support ticket exploiting the agent's own legitimate access. A separate incident involving Drift and Salesforce showed the same dynamic, but at larger scale. Attackers identified as UNC6395 exploited OAuth tokens stolen through phishing and technical exploitation of Salesloft's infrastructure to gain unauthorized access to hundreds of corporate Salesforce environments and exfiltrate sensitive data. OX Security's research, titled "The Mother of All AI Supply Chains," found vast numbers of vulnerable instances spread across a supply chain tied to hundreds of millions of package downloads.
None of this trips the alarms security teams already have. Most MCP attacks travel over legitimate HTTPS traffic and use valid tool invocations, and WAFs, DLP tools, SIEMs, and API gateways were never built to evaluate the intent behind an agent's tool call. Wiz Research describes the underlying danger as a "lethal trifecta": untrusted input, access to sensitive data, and the ability to act, a combination that static controls simply can't address. The NSA has flagged risks specific to MCP systems: dynamic tool invocation, where agents call new tools autonomously at runtime, and implicit trust relationships, where one agent's output gets treated as valid by another agent without any verification step. Security teams are reacting to a named, recorded pattern of incidents that their existing tools are structurally unequipped to see.
The identity problem at the root of most agent security failures

Across most of these failures, the root cause is the absence of a proper identity architecture for the agents themselves, not a software bug or a weakness in the underlying model. Machine identities already outnumber human identities at enterprise organizations by at least 45 to 1, and that ratio is accelerating as AI adoption spreads. Without formal registration, scoped permissions, an assigned owner, and inclusion in audit logs, agents accumulate as shadow identities, which makes them the most capable actors in the infrastructure and the least overseen.
The access design itself compounds the problem. Agents determine what they need at runtime, dynamically, so any permission set provisioned ahead of time either over-grants access and expands the blast radius, or under-grants it and causes failures that get resolved by handing over even broader access. Agents can also be manipulated through prompt injection, so they reveal environment variables, exfiltrate credentials, or escalate their own permissions, and that turns the agent itself into an attack surface that no static credential control can address.
Fixing this starts with attribution. Every agent needs its own distinct identity, never a shared service account and never a borrowed human credential, because attribution is the precondition for every control that comes after it: audit evidence that can't separate agent activity from human activity can't support compliance. Runlayer, an enterprise AI agent control plane that enforces policy at the tool-call level and ties a verified identity to every agent session across 18,000+ integrations, is built around closing exactly that gap before agents ever reach production. Credential design should lean on workload identity federation and short-lived, automatically rotated credentials rather than secrets embedded in code. Where an agent is acting on a user's behalf, OAuth 2.0 Token Exchange, specified in RFC 8693, preserves the distinction between the agent's own identity and the authority it's been lent. Agents should never hold long-lived credentials at all; issuing short-lived credentials at runtime, scoped to the specific task, shrinks the exposure window to minutes even if the agent is compromised.
The industry is converging on this model. Microsoft Entra Agent ID, generally available since April 2026, gives each AI agent its own identity for lifecycle and access management, with full integration into Conditional Access and Microsoft Purview available through Microsoft Agent 365, the broader control plane sitting on top of it; third-party agents built on AWS Bedrock, n8n, and similar frameworks can be onboarded through workload identity federation. At RSA Conference 2026, Cisco announced security innovations for the agentic AI ecosystem, including registering agents in Duo IAM and mapping each one to an accountable human owner, so every agent carries a verified identity and its actions stay traceable. McKinsey's "Lilli" red-team exercise named the cause this way: an autonomous agent compromised the internal AI platform, gaining broad system access in under two hours, faster than human responders could contain it.
Why governance gaps stall pilots before they reach production
Organizations can run agents without governance controls as a neutral default, but they can't sit in that indefinitely. It's the single largest reason AI agent pilots never reach production, and for the organizations that push forward anyway, it builds regulatory and business exposure that compounds over time. An estimated 11 to 14 percent of enterprise agentic AI pilots make it to production; the rest stall on gaps in identity, audit, and access control, not on whether the model works or the use case holds up. Cisco's survey of major enterprise customers found the same bottleneck from the other direction: a large majority of organizations reported experimenting with AI agents, but only a small fraction had moved that technology into production, with the security gate standing as the largest single obstacle between experimentation and real value.
The regulatory clock is no longer abstract. The Digital Omnibus on AI, signed 8 July 2026 and in force since 27 July 2026, made the EU AI Act's high-risk obligations enforceable, with stand-alone high-risk systems bound starting December 2, 2027, and AI embedded in regulated products bound starting August 2, 2028. Those deadlines cover AI used in employment, education, critical infrastructure, and financial services. Most enterprise agentic deployments already sit in those areas. The EU AI Act binds deployers even when they didn't build the underlying model themselves, and it can fine them a substantial percentage of global revenue. Gartner projects that by 2028, a significant share of enterprise generative AI applications will see five or more minor security incidents a year, up sharply from 2025 levels, and it names MCP as a key contributing factor alongside immature security practices and the broader spread of agentic AI.
Inside the organization, the people responsible for managing this risk are themselves caught in a bind. Fewer than a third of CISOs feel fully aligned with their C-suite and board on what level of AI risk is acceptable, and less than half say their board treats AI security as a business enabler rather than a compliance checkbox. That misalignment leaves security teams asked to approve deployments they have no way to govern, while also taking the blame for slowing adoption down, a position where blocking rollouts is the rational, risk-minimizing move available to them. The MCP specification itself already lays out what production use requires: OAuth 2.1 with PKCE for every protected HTTP-based deployment, HTTPS on all endpoints, discoverable authorization server metadata, audit trails, and SSO-integrated authentication. Only a small fraction of MCP implementations use OAuth in practice, and that gap between what the spec demands and what gets deployed is what stalls pilots before they reach production.

Giving security teams real control in practice
The path forward is building the specific controls that let security teams keep their bar in place while adoption moves forward anyway.
Visibility has to start when you build a complete, real-time inventory of every MCP server, agent skill, and tool in use across the organization. That inventory can't be a spreadsheet maintained by hand. It needs to be a live registry that shows what's connected, who owns it, and what permissions it holds at any given moment. On top of that inventory, identity needs to bind to every action at the level of individual tool calls: single sign-on and SCIM tying agent activity back to human identities, just-in-time credentials that expire the moment the task ends, and OTEL tracing that produces a tamper-proof chain of custody running from a user's original intent through every tool the agent invoked along the way. Enforcement comes next, and it has to work at tool-call depth rather than at the perimeter: instead of one blanket rule approving an agent across the board, policy needs to specify which tools a given agent can invoke, under what conditions, and with what level of human approval required for higher-risk actions. Detection has to match that same granularity, catching MCP-specific threats like tool poisoning, prompt injection, intent drift, and credential exfiltration by reading the semantic content of a tool call rather than scanning for the network signatures that traditional security tools were built to catch.
Regulators are already moving to formalize this model, so they don't leave it to individual vendors. Singapore's Infocomm Media Development Authority launched the first governance framework written specifically for autonomous AI, and that signals agentic AI needs its own governance model, built for how it actually behaves, not an extension of governance built for conventional software. The operational shift this adds up to is straightforward: security teams review and approve agents against a defined policy before deployment, and that review isn't a one-time gate but a continuous process, backed by audit logs that make every agent's access and action traceable long after the approval was granted. Once that foundation exists, security no longer stands as the obstacle between a pilot and production. It is the mechanism that makes production possible.


