• 9 min read

46% of Teams Hit by AI Agent Incidents: MCP Audit Checklist

tl;dr

Shadow MCP is the critical blind spot your gateway misses. Gateways govern only their own traffic while unapproved MCP servers run as local JSON files on laptops with live credentials and no audit trail.

Featured image for "46% of Teams Hit by AI Agent Incidents: MCP Audit Checklist"

Nearly half of surveyed organizations reported an AI agent security incident or near-miss in the past year, according to a VentureBeat Intelligence survey of 102 AI practitioners. That number should stop you cold — not because it’s surprising, but because it confirms what many engineering leaders already suspect: agents are operating with more access than anyone tracked, and the incidents are piling up faster than governance can catch up.

What Does the 46% Incident Rate Actually Mean for Your Team?

The VentureBeat Pulse report breaks down the 46% into 6% confirmed incidents and 40% near-misses caught before harm. But the near-miss figure is arguably more telling — it means teams are discovering problems only because someone happened to notice, not because a control fired. Manifold Security’s Agent Incident Ledger lists 59 publicly disclosed incidents since July 2025, and only 11 were identified by a security control or safeguard. The rest surfaced through later reviews, outside reports, or after damage was already observed.

ServiceNow’s research paints an even starker picture: 90% of organizations worldwide experienced an agentic AI security incident in the past year, yet only 13% have a defensible security posture. IBM’s 2026 Tech Leader Study found enterprises detected an average of 54 AI agent incidents over 12 months — 37% resulted in data exposure, 33% triggered cascading system failures, and 17% created compliance issues. The pattern is consistent: adoption is outpacing visibility, and the incidents that do get caught are caught late.

When an incident does happen, the response gap is measurable. Manifold found a median of 70 days between an agent’s first unauthorized action and public disclosure, with a median of 20 days after discovery before disclosure. Incidents detected by security controls actually took longer to become public — 111 days median versus 50 days for other detection methods. That suggests controls are catching things, but the triage and disclosure process is slow. Meanwhile, Microsoft Threat Intelligence observed an AI agent advance from scanning public GitHub repositories for exposed secrets to exfiltrating data from Azure Key Vault in just five minutes. The attack surface moves at machine speed; your governance moves at meeting speed.

Why Shadow MCP Is the Blind Spot Your Gateway Misses

Shadow MCP refers to unapproved Model Context Protocol servers that employees wire into AI clients like Claude Desktop, Cursor, or Claude Code without IT review — each holding a live token to a corporate system and existing as a local JSON configuration file on a laptop. Strac’s analysis describes it as the agent-era version of shadow IT, but worse: there’s no OAuth consent screen, no admin console, and no audit trail. The configuration is a local JSON file on a laptop.

Cloud MCP gateways cannot see shadow MCP servers because they only govern traffic that flows through them, and an employee can point an AI client directly at any server to bypass the gateway. Strac’s blog makes this explicit: a gateway governs only its own traffic, and pointing a client straight at a server bypasses it entirely. The only reliable place to see shadow MCP is on the endpoint, because the configuration lives in a local JSON file on the employee’s laptop with no OAuth consent screen, admin console, or audit trail.

This isn’t theoretical. Salt Security’s survey of 380+ professionals found 32.9% of respondents had found an unauthorized, shadow, or misconfigured MCP server in the past 12 months, and a wider 54.3% had found or suspected one. Salt Labs detected MCP-related infrastructure in 21.6% of customer environments in August 2026, up from 16.7% in January. And it’s not always employee-initiated — a developer reported that opening Postman’s desktop app automatically wired a Postman MCP server into both Claude Code and OpenAI’s Codex without asking. The sequence was marked “auto-existing-upgrade” in Postman’s own logs. Shadow MCP can arrive without anyone choosing to install it.

Reco’s analysis of 500 MCP servers found 50% enable shell command execution, 82% enable local file reads and writes, 73% support outbound network calls, and 40% allow the full trifecta of all three capabilities. That’s a remotely reachable tool with host-level reach and no lock on the door for about 1 in 8 of these tools. The same report found 80% of AI tools used by employees — including browser extensions and MCP servers — were ungoverned, compared with 79% of SaaS applications that were authorized. Shadow AI is outpacing shadow IT at many organizations.

What the MCP Server Data Reveals About Credential Hygiene

A survey of 5,200 open-source MCP servers found 88% require credentials, 53% rely on long-lived static secrets, only 8.5% use OAuth, and 79% pass API keys through environment variables. Alyria’s analysis cites the Astrix “State of MCP Server Security 2025” for these numbers, and they’re alarming: durable secrets on laptops in files no security tool reads. That’s a configuration an employee adopts in an afternoon, and it puts a live credential on an endpoint with zero governance.

The credential problem compounds when agents share identities. VentureBeat’s survey found 57% of enterprises say each AI agent in production has its own scoped, managed identity, but 39% say some or most agents share credentials. Among enterprises giving each agent its own identity, 60% say a named owner could revoke access within minutes — against 33% of enterprises whose agents share credentials. Revocation speed matters: 49% of surveyed enterprises say a named owner could revoke an agent’s access within minutes, 36% say a security or platform team could do it within hours, and 12% say it would take more than a day or require tracing the credential first.

IBM’s research illustrates the failure mode: a user asks an innocent question, the agent selects a tool, the tool queries a system under a different identity than the user’s, and the response returns data the user’s identity shouldn’t have unlocked. Every component did exactly what it was supposed to in isolation, yet the system violated the intended access model. The abstraction adds another layer to the trace — now organizations must ask which agents can call which servers, what identity each server presents, and whether the user’s authorization holds across the full chain.

How Do You Audit Repo Tokens and Agent Permissions?

Start with the configuration files on developer endpoints. AI coding assistant configuration files — settings.json hooks, .cursorrules, Cursor MDC rules, and similar harness files — are now explicit targets in supply-chain attacks. Tenable’s research found the Mini Shai-Hulud worm scans the developer’s home directory for configuration files belonging to AI coding tools by name and injects hook commands for silent persistence. Every time the developer starts an AI coding session, the malware runs with the same trust level the developer placed in their own tools.

These harness files sit at the intersection of three trust relationships: the developer trusts them as config, the IDE executes them automatically, and the LLM treats them as authoritative instructions. Tenable recommends treating harness config files as code that requires mandatory reviews and hash pinning in the CI/CD pipeline, enforcing —ignore-scripts on package installs, and flagging AI scanner refusals as a suspicious signal rather than a clean result.

For repo tokens specifically, audit what each agent can actually reach. GitGuardian’s data shows Claude Code-assisted commits introduced secrets at 3.2%, double the 1.5% baseline. The practical controls: run the agent in a sandbox, give each agent a credential scoped to its task, keep configuration and MCP servers out of reach, record every call, scan for secrets, and deny by default. Novee Security’s Black Hat research found a public GitHub issue opened by an outside user with no repository privileges could turn AI coding tools against the repositories they were meant to maintain — paths to remote code execution, credential theft, and repository modification existed in default workflow configurations.

Which Governance Tools Actually Publish Pricing?

ToolPricing ModelFree TierTarget Audience
RunlayerCustom quote only; no published pricing, no self-serve signup, no free trialNoneEnterprise teams needing managed MCP infrastructure and agent governance
Willow15 users free forever; Enterprise priced per human seat, billed annually, on request15 usersTeams wanting governed MCP gateway with identity layer and risk scoring
TrueFoundryPublished MCP Gateway with RBAC; pricing visibleNot specified in researchPlatform teams wanting self-serve MCP governance without custom enterprise deals

Runlayer does not publish pricing or plan tiers; every deal requires a custom quote after contacting sales, with no self-serve signup, no starts-at number, and no publicly advertised free trial. TrueFoundry’s breakdown notes the quote likely depends on organization size, number of users, deployment model, enterprise requirements, and support level — but you can’t sanity-check a number until you’re already in a sales cycle. Willow offers 15 users free with $0 forever pricing; Enterprise is priced per human seat, billed annually, on request. The Willow review notes 1,000+ pre-built connectors, 0-10 risk scoring per MCP server, and five deployment options on Enterprise including air-gapped.

A shadow MCP remediation or governance cost projection for an engineering team cannot be constructed from the research: Runlayer does not publish pricing, Willow Enterprise pricing is quoted per seat on request, and no per-incident or per-team remediation cost data is provided. This gap means you’ll need to run your own vendor evaluations with real scope before budgeting. VentureBeat found 79% of organizations that reported an AI agent incident or near-miss plan to adopt a new, additional, or replacement agent security solution within 12 months, compared with 47% of those reporting no incident or near-miss. The buying signal is clear — but the pricing transparency isn’t there yet.

What Should Your Checklist Cover Before Granting Write Access?

Inventory the whole agent stack first. Most teams already track npm, Python, and container dependencies. Add MCP servers and their package versions, agent skills and prompt packs, editor extensions, lifecycle hooks and shell scripts, local tools the agent can invoke, and remote services the agent can authenticate to. Keep the inventory in the repository or an internal document. Record the source, version or commit, maintainer, permissions, and why the team needs it. If you cannot answer “what code runs when this agent starts?”, your inventory is incomplete.

Review capabilities, not marketing descriptions. A tool described as “GitHub integration” might read issues, create branches, push commits, open pull requests, or access repository secrets — very different risk levels. For every extension or MCP server, ask: Can it read files outside the current project? Can it execute shell commands or start processes? Can it access environment variables, tokens, or SSH credentials? Can it make outbound network requests? Can it modify files or publish changes without approval? Write down the actual capabilities. Do not infer them from the tool’s name or marketplace description.

Apply least privilege at the process boundary. Agent approval prompts are useful but not a complete sandbox for code running alongside the agent. Run experiments in a disposable container or VM. Mount only the repository and directories the task needs. Keep production credentials out of local agent environments. Use short-lived, narrowly scoped tokens. Restrict outbound network access where practical. Require human approval for external side effects. A useful default: an agent working on a feature branch should not inherit credentials that can deploy production.

Pin, verify, and update deliberately. For every component, prefer a specific version or immutable commit over a floating branch or an unpinned “latest” install. But do not confuse pinning with freezing — schedule regular reviews to update pinned versions after verifying changelogs and running tests. Treat agent extensions like production dependencies, because that’s what they’ve become. For a deeper dive on auditing the full agent supply chain, see how to audit your AI coding agent’s MCP supply chain. If you’re dealing with shadow agents already writing code from developer endpoints, endpoint detection is required before agents merge pull requests.

The 46% incident rate isn’t a forecast — it’s a baseline. Your agents are already running with credentials you haven’t fully mapped, connecting to MCP servers you haven’t inventoried, and writing to repositories you haven’t scoped. The checklist above isn’t optional hygiene; it’s the minimum viable control plane for any team running agents in 2026.