• 9 min read

Shadow AI Detection: Stop Unauthorized Tools and Code Leaks

tl;dr

Network security tools miss shadow AI because the risk has moved to developer endpoints. Coding agents and local MCP servers run with full developer permissions outside AI gateways and network controls. Engineering teams need endpoint and developer-layer detection, not just CASBs and secure web gateways.

Featured image for "Shadow AI Detection: Stop Unauthorized Tools and Code Leaks"

Your security team approved 300 AI agents. CrowdStrike found 18,000 running on endpoints. That gap isn’t an edge case — it reflects what one Fortune 500 customer discovered when they turned on agent discovery, and it illustrates the scale of the visibility problem engineering organizations face in 2026.

Why Network Tools Miss the Real Shadow AI Risk

The shadow AI problem has moved. It used to live in browser tabs — ChatGPT, Claude, Gemini — where network-layer tools like CASBs and secure web gateways could see the traffic. Today the highest-risk AI runs on developer laptops as coding assistants, local MCP servers, and agent skills that never make an external AI-domain network call. A Fortune 500 customer using CrowdStrike Falcon Guardian discovered 18,000 active AI agents on endpoints despite having approved only 300. Claude Code, OpenAI Codex, Cursor, and Kiro were among the agents Guardian identified.

This is what I call the Endpoint Shadow Gap: the largest blind spot now sits exactly where AI execution and data access happen — on developer machines. Network-plane detection sees managed device traffic but misses local agent, MCP stdio, and localhost connections. Endpoint-plane detection sees local agents and MCP servers but requires agent deployment and misses pure network-layer context. Most organizations have invested heavily in the first plane while the risk has migrated to the second.

90% of developers use AI coding agents at work at least weekly, creating endpoint security gaps where agents can read files, run shell commands, and call MCP tools directly on that developer’s machine. These agents operate with the developer’s permissions, outside every AI gateway and most network controls. The control point has moved to the agent’s hooks — pre-tool callbacks that can allow, ask, or deny an action before it runs — but most security tools still watch the network.

What the Data Shows About Shadow AI Density

Scale doesn’t shrink this problem the way security teams assume. Organizations with 11 to 50 employees average 269 unsanctioned AI tools per 1,000 employees, while mid-sized organizations of 500 to 1,000 employees average roughly 200 per 1,000. The density barely declines with company size. Small firms don’t have a monopoly on shadow AI — mid-sized firms are only marginally better.

Risk concentration and tool sprawl are simultaneous patterns. OpenAI separately accounts for 53% of all shadow AI usage measured in Reco’s 2025 State of Shadow AI Report, meaning a single vendor dominates the exposure surface even as the long tail of tools keeps growing. Employees adopting unauthorized AI tools are rarely malicious; most are trying to boost productivity with whichever tool solves today’s task fastest. Outright bans push the behavior further out of view rather than closing it.

One in five organizations reported a breach due to shadow AI, and among those breached through an AI model or application, 97% lacked proper AI access controls. Tenable detected 457 million AI-related security issues across more than 7,000 organizations, averaging 62,000 exposures per organization over a 30-day period. The exposure surface is massive and growing.

How Detection Planes Compare for Engineering Teams

Shadow AI detection splits into three planes — network, browser/endpoint, and identity/spend — and a program covering only one plane has a blind spot the size of the other two. Here’s how they map to engineering reality:

Network plane (Netskope, Cloudflare, Microsoft Global Secure Access): Sees traffic from managed devices. Netskope’s inline inspection covers 1,800+ AI apps, distinguishes corporate from personal accounts, and can coach, redact, or block on a custom quote. Cloudflare is the only one of ten shadow AI detection tools reviewed that offers a free plan; the other nine are paid or contact-sales. Microsoft Purview costs $12 per user per month for Microsoft 365 shops over E3 licenses. Network tools catch browser-based ChatGPT sessions but miss local coding agents and MCP stdio connections entirely.

Browser/endpoint plane (Harmonic Security, Prompt Security, Nightfall AI, LayerX, AgileBlue): Sees the prompt and the account on the device. Nightfall AI delivers approximately 95% detection precision out of the box against a 5% to 25% accuracy baseline for legacy DLP, and cuts false positives by 99%. AgileBlue Shadow AI Monitoring Phase 2 detects local and cloud-hosted MCP servers within 15 minutes of activation, including local servers that do not appear in any configuration file. These tools catch what network tools miss but require endpoint agent deployment.

Identity/spend plane (Reco, Torii, Zylo): Sees SSO logins, OAuth grants, and card charges. This is where AI built into sanctioned SaaS and connected agents show up — the AI feature switched on inside a CRM, or the agent a team connected with an OAuth grant. It misses local-only tools that don’t use corporate identity.

Developer/MCP layer (JFrog, Obot, Pillar, DeepKeep, C1.ai): The newest layer, purpose-built for coding agents, skills, plugins, and MCP servers on developer workstations. General shadow AI tools treat this as a footnote or skip it outright. JFrog’s AI Security Agent Plugin, deployed via MDM or agent harness settings, detects skills, MCP servers, plugins, hooks, and monitors on developer workstations and reports metadata and risk analysis to the JFrog Platform. JFrog’s 2026 Software Supply Chain Security report identified 969 malicious agent skills with high-impact payloads and 495 malicious AI models on Hugging Face. This layer is where the risk concentrates for engineering teams.

What Tools Actually Catch Coding Agents and MCP Servers

If you’re securing an engineering organization, you need visibility into the developer/MCP layer specifically. Here’s what the current crop of tools covers:

JFrog AI Security extends shadow AI detection to developer workstations via a plugin deployed through MDM or agent harness settings. It inventories skills, MCP servers, plugins, hooks, and monitors, runs deterministic rules-based scans locally, and reports metadata and risk analysis to the JFrog Platform. The BYOK semantic scanning option lets regulated teams use their own LLM endpoints for asset analysis. JFrog also added Agent Guard for policy enforcement at the point an agent reaches for a skill or model.

AgileBlue Phase 2 detects AI coding assistants (Copilot, Cody, Tabnine, and similar tools including IDE extensions in VS Code, JetBrains, Cursor), AI browser extensions across Chrome, Edge, Firefox, Arc, and Brave, AI files and artifacts (model files, Cursor rules, Claude Desktop configs, plugin and project files), and unauthorized MCP connections — both local and cloud-hosted — within 15 minutes of activation. Detections include user and process attribution.

Nightfall AI provides unified coverage across endpoints, MCP servers, email, browsers, and SaaS in one platform. Its endpoint agent deploys via MDM in about 30 minutes with full inline blocking rather than alerts alone. The AI-based data lineage tracks sensitive data from source to destination through downloads, renames, and compression. Nightfall prices per user per year with AI capabilities native in every tier.

CrowdStrike Falcon Guardian leverages the existing Falcon sensor for agent discovery. The capability shipped as a Falcon Flex entitlement, meaning existing customers could activate it without a separate procurement cycle. Guardian identified Claude Code, OpenAI Codex, Cursor, and Kiro among the 18,000 agents at that Fortune 500 customer. The single-agent architecture means no new endpoint footprint.

GitHub code scanning for AI applications (public preview) detects indirect prompt injection, sensitive information disclosure, and excessive agency in pull requests. It requires a GitHub Advanced Security license and runs automatically on PRs alongside CodeQL analysis. This catches AI risk at the code-review stage but doesn’t see runtime agent behavior on developer machines. GitHub code scanning for AI applications, in public preview, detects indirect prompt injection, sensitive information disclosure, and excessive agency in pull requests, but requires a GitHub Advanced Security license.

DeepKeep AI Lens for Developers deploys as a lightweight plug-in rather than a full endpoint agent. Hooks built into the coding agent intercept prompts, shell commands, file reads, and MCP tool calls before and after they run, routing each to DeepKeep for an allow, block, or audit decision. It flags credentials, tokens, passwords, and insecure code patterns, with destructive commands sent for human approval before execution.

How Much Does Shadow AI Detection Cost for a 50-Person Team

Pricing varies wildly by plane and vendor. For a 50-engineer team, Microsoft Purview costs approximately $7,200 per year in Purview licensing alone (50 engineers × $12/user/month × 12 months), excluding any required Microsoft 365 E3 base licenses. Netskope operates on custom quotes. Cloudflare offers a free tier. Most endpoint and developer-layer tools price per user per year or per seat with volume discounts — but public pricing is rare. A 50-engineer team faces substantial base AI coding subscription costs before metered usage overages, so detection costs stack on top of existing AI tool spend.

ToolDetection PlanePricing ModelBest For
Microsoft PurviewNetwork (M365/Edge)$12/user/month over E3Microsoft 365 shops needing identity + network visibility
NetskopeNetwork (inline)Custom quoteMixed estates needing 1,800+ app coverage with coach/redact/block
CloudflareNetwork (CASB)Free plan availableTeams wanting zero-cost network-layer visibility to start
Nightfall AIEndpoint + MCP + SaaSPer user/year (AI native in all tiers)Unified data movement control across humans and agents
AgileBlueEndpoint + MCPContact salesSOC teams already on AgileBlue wanting MCP detection in 15 min
JFrog AI SecurityDeveloper/MCP layerPart of Unified/Ultimate SecuritySoftware supply chain teams needing skill/model governance
CrowdStrike Falcon GuardianEndpoint (agent)Falcon Flex consumptionExisting Falcon customers wanting zero-new-agent deployment
GitHub Code Scanning (AI)Code review (PR)Requires GitHub Advanced SecurityTeams already on GHAS wanting AI-specific PR-time detection

The 50-engineer shadow AI detection deployment using Microsoft Purview illustrates the math: base licensing alone hits $7,200/year before any E3 costs. Add endpoint-layer tools and the total climbs fast. Most vendors don’t publish per-seat pricing, so you’ll need quotes. Budget for at least two planes — network plus endpoint/developer — because one plane leaves a blind spot the size of the other two.

What a Layered Detection Strategy Looks Like in Practice

Start with discovery, not enforcement. An observe-only phase builds inventory and dashboards without breaking workflows — useful for a first audit, but it leaves remediation to other systems. The strongest programs pair discovery with the ability to act: redact a sensitive value, block a file heading to an unmanaged AI tool, or warn the user in the moment. Ideally in one agent so there’s one policy and one audit trail.

For engineering teams, the minimum viable coverage is network plane (for browser-based AI) plus developer/MCP plane (for coding agents and local MCP servers). If you’re a Microsoft 365 shop, Purview gives you the network plane at known cost. Pair it with JFrog, AgileBlue, Nightfall, or CrowdStrike Guardian for the endpoint/developer layer. If you’re not Microsoft-centric, Netskope or Cloudflare covers network, and you still need the developer layer. Endpoint agents are required to catch shadow AI in engineering organizations — network and browser tools miss CLI, IDE, and MCP traffic where sensitive code and credentials actually flow.

Privacy-preserving metadata-only detection avoids surveillance concerns but cannot see what data actually reached the AI tool. Deep content inspection of prompts and files catches sensitive data leakage but requires keystroke, file, or screen access. Nightfall and Harmonic scan prompts on-device with configurable retention; AgileBlue and JFrog report metadata and risk analysis. Choose based on your regulatory environment and developer trust requirements.

Don’t evaluate tools by dashboard features. Buy by detection plane. Because most enterprise AI risk in 2026 now resides on developer endpoints and in local MCP connections rather than in browser-based ChatGPT sessions, network-first CASB and secure-browser investments will leave the largest blind spot unaddressed unless paired with endpoint agent and coding-agent controls. The Shadow AI compliance requires discovery before documentation — and that discovery must cover the plane where your developers actually work.

What’s your current coverage gap — network, endpoint, or the developer/MCP layer?