• 6 min read

Shadow AI in Engineering: Detect and Govern Coding Tools

tl;dr

Shadow AI in engineering is a visibility problem, not a policy problem. The most dangerous tool is a sanctioned agent with disabled controls, and organizations with high shadow AI see average breach costs of $670,000.

Featured image for "Shadow AI in Engineering: Detect and Govern Coding Tools"

Seventy-eight percent of enterprises report employees using unauthorized AI tools for coding tasks, yet ninety-two percent of Fortune 500 companies have deployed at least one sanctioned AI coding assistant. That gap between what leadership approves and what developers actually use isn’t a policy problem — it’s a visibility problem. When a developer pastes a function containing hardcoded credentials into a personal Claude.ai account, that code leaves your environment under Anthropic’s consumer data policy, not your enterprise agreement. The same paste into GitHub Copilot Individual routes through Microsoft’s consumer infrastructure with different retention terms. Most developers don’t know the difference; they see the same chat interface and assume enterprise controls apply everywhere.

How big is the shadow AI problem in engineering?

The numbers are stark. Eighty-one percent of employees and eighty-eight percent of security leaders use unapproved AI tools, and forty-five percent of workers find a workaround when their employer blocks an app, according to UpGuard’s shadow AI research. Blocking alone doesn’t eliminate demand — it just moves it somewhere you can’t see. Ninety-four percent of organizations are using or piloting GenAI coding tools, and forty-eight percent said the tools have increased their use of open source software, expanding the supply chain surface area that traditional SBOMs were never designed to track. Meanwhile, seventy percent of teams expect AI components in production by the end of 2026, yet forty-three percent have no formal governance over which components developers can use. The scale surprises most security leaders the first time they measure it — discovery exercises routinely turn up assistants, extensions, and model integrations across teams that had reported using nothing at all.

What does shadow AI actually cost you?

Organizations with high levels of shadow AI observed an average of $670,000 in higher breach costs than those with low or no shadow AI. JetBrains’ AI spend grew roughly tenfold in the first half of 2026, and its first cross-vendor usage report took four days to compile by hand — for a team of roughly 2,300 people, per JetBrains’ Air Governance announcement. That’s the hidden tax: not just the subscription fees, but the operational overhead of stitching together visibility across Claude Code, Cursor, Copilot, Codex, and whatever tool a developer installed last week. A 50-developer team governing AI coding agents and tools with Olakai Agentic spends $15,000 per year in subscription fees [50 developers × $25/mo/developer × 12 months].

ToolTierPriceCoverage
Olakai AgenticPer developer$25/monthBudgets, forecasts, cost-per-merged-PR, model routing for Claude Code, Cursor, Copilot, Codex
Olakai AssistivePer employee$5/monthShadow AI detection across 800+ tools, DLP, adoption tracking by team/role/region
Microsoft PurviewPer user (E5)$12/monthNative M365 and Copilot controls
GitHub CopilotBusiness$19/user/monthPolicy management, audit logs
CursorTeams$30/user/monthSSO, access controls on Enterprise
Claude CodeEnterprise$20/seat/month annual ($25 monthly)Enterprise plan, VPC via Bedrock/Vertex/Foundry

The pricing dispersion is real — and it’s before metered usage overages. You’re not buying a single tool; you’re managing a portfolio where each vendor bills differently, each dashboard shows different metrics, and none of them talk to each other.

Why do blocking strategies fail?

Forty-five percent of workers find a workaround when their employer blocks an app. Only seventeen percent of organizations have technical controls that block access to public AI tools combined with DLP scanning, per the Kiteworks survey cited by Torii. The rest rely on policy documents that developers don’t read and network controls that miss browser-only traffic. Adversa AI’s GuardFall research (June 2026) bypassed the command-approval guards in ten of eleven popular open-source coding agents with decades-old shell tricks, and auto-approve flags reopened the hole even in the one agent that held. Google’s April 2026 Gemini CLI advisory patched remote-code-execution paths whose preconditions were --yolo mode and auto-trusted workspace folders. The agents’ built-in guards are not a control you can rely on — they’re features the vendor can change or that a determined developer can disable.

How do you detect what you can’t see?

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. An endpoint sensor can tell you Claude Code is installed. It cannot tell you that it runs with permission prompts disabled, auto-approves a community MCP server, and loads a skill that executes shell before the model reads it. Those are configuration facts, and they live in the agent’s own settings files. JetBrains built Air Governance to solve exactly this — one console to manage which AI agents, models, and providers teams use, set spending limits, and see what that usage costs and delivers across IDEs, CLI, web browsers, and mobile. It works with Claude Agent, Codex, Gemini CLI, Junie, and other ACP-compatible agents. The control point has moved to the agent’s hooks: Claude Code, Codex CLI, Cursor, Gemini CLI and most newer harnesses expose a pre-tool hook that can allow, ask, or deny an action before it runs. In 2026 Microsoft, CrowdStrike, Palo Alto Networks, Noma, Zenity, Straiker and two open-source projects all adopted it. The question is no longer whether a tool uses hooks, but which agents, which actions, and what happens when the hook is removed or times out.

What governance framework actually works?

Forty-three percent of companies still lack a formal AI risk management framework, according to Gallagher’s survey of more than 1,200 companies. The EU AI Act’s Article 50 transparency obligations took effect on August 2, 2026, and prohibited practices under Article 5 carry penalties of up to 35 million euros or 7 percent of worldwide annual turnover. Eighty percent of recorded AI agent incidents lacked any public documentation showing security controls were in place or functioning at the time, per Anaconda’s findings. That documentation gap means most incident post-mortems cannot produce the control-efficacy evidence that insurers, auditors, and boards increasingly require. Gaps in incident records become gaps in insurance coverage and audit findings. The framework needs to shift from tool discovery to agent configuration and authoring context — because endpoint sensors and artifact gates can confirm an AI agent is present or its code is clean, but cannot see whether safety hooks are disabled, MCP servers are auto-trusted, or what decisions the agent made while writing the code. The most dangerous enterprise AI is not an unsanctioned shadow tool but a sanctioned agent with disabled controls, because governance dashboards report it as approved while its actual configuration — permission prompts off, auto-approve flags set, hooks failing open — remains invisible to discovery scanners.

Which tools should you evaluate?

Start with the control surface that matches your risk. If your primary exposure is employee chat usage across 800+ tools, Olakai Assistive at $5/month per employee gives you shadow AI detection, DLP, and adoption tracking by team, role, and region, per Olakai’s pricing page. If you’re managing coding agents and need budgets, forecasts, and cost-per-merged-PR tracking across Claude Code, Cursor, Copilot, and Codex, Olakai Agentic at $25/month per developer covers that layer, per Olakai’s pricing page. Microsoft Purview at $12/user/month (E5) makes sense if you’re already deep in M365 and Copilot. But no single tool covers the full stack — most enterprises end up pairing two. The evaluation criteria should be: does it enforce at the agent’s hook layer or only report at the artifact gate? Does it fail closed when the daemon is unreachable? Does it preserve the authoring trail — retries, tool calls, context gathering — or only the finished diff? Artifact gates evaluate only the finished code at the decision boundary, where agent-written code looks identical to hand-written code. The system running the agent holds the full record. That’s where governance has to live.

What’s your team’s actual control surface today — and which agent configurations would your current tooling miss?