Shadow AI is any AI tool, model, feature or integration used for work without formal approval or oversight. It is the successor to shadow IT, with one important difference: shadow IT stored your data somewhere you did not know about, whereas shadow AI reads your data, may retain it, and increasingly acts on your systems. The six risks follow from that — loss of visibility, data and privacy leakage, regulatory exposure, a wider attack surface, unreliable or biased output, and tool sprawl with vendor lock-in.
1 · The Eight Hiding Places
Roughly in order of how visible they are to a security team — the first is easy to spot, the last is nearly invisible.
| # | Where it hides | Typical use | Main risk |
|---|---|---|---|
| 01 | Employee browsers — public chatbots, AI browser extensions | Drafting emails, summarising documents, analysing pasted data | Sensitive text leaves approved systems; extensions can read every page |
| 02 | Personal accounts — privately registered AI accounts used for work | Content, research, code, analysis | No enterprise contract, no retention terms, no offboarding |
| 03 | SaaS applications — AI features switched on inside approved tools | Meeting notes, recommendations, document generation | A new data processor arrives through an update and skips vendor review |
| 04 | Department tools — AI platforms bought directly by a team | Marketing, sales intelligence, customer analytics | Duplicate spend, unreviewed contracts, compliance gaps |
| 05 | Developer environments — unapproved models, APIs, coding assistants, libraries | Code completion, tests, refactoring | Proprietary code and secrets reach external providers |
| 06 | Workflow integrations — AI wired in through APIs, plugins, automation platforms | Record updates, data transfers, routing | Over-broad OAuth scopes give an AI service standing access |
| 07 | Uploaded documents — enterprise files pushed into external AI | Contract review, translation, extraction | Confidential files stored or reused outside your control |
| 08 | Embedded AI agents — third-party agents operating inside your applications | Task execution, orchestration, support automation | Software that acts with no owner, no monitoring and no approval |
▸ Places 06 and 08 are where shadow AI stops being a data-leak problem and becomes an access-control problem: the AI is no longer just reading what someone pastes, it holds credentials and takes actions.
2 · How to Find Each One
Every hiding place leaves a trace in a system you probably already run. Discovery is mostly about reading logs you already collect with a new question in mind.
| Hiding place | Where the evidence is | Control once found |
|---|---|---|
| Browsers | Secure web gateway / DNS logs for AI domains; managed-browser extension inventory | Allow-list approved AI; block unvetted extensions; DLP on paste/upload |
| Personal accounts | Gateway logs showing consumer (non-SSO) logins to AI services | Enterprise tenant with SSO; tenant restrictions to block personal sign-in |
| SaaS features | SaaS admin consoles, release notes, SaaS posture management | Default AI features off until reviewed; add AI clauses to vendor review |
| Department tools | Expense claims, card spend, procurement records | AI purchases routed through one intake and risk review |
| Developer environments | Repository scanning for AI SDK imports and API keys; egress logs to model APIs | Approved assistant with no-training terms; secret scanning; egress via the AI gateway |
| Integrations | Identity provider OAuth-grant and app-consent reports | Admin consent for AI apps; least-privilege scopes; periodic grant review |
| Uploaded documents | DLP and gateway logs of file uploads to AI domains | Classification labels that block upload of confidential files |
| Embedded agents | Service-account and API-key inventories; unusual machine identities | Each agent gets its own identity, an owner, scoped tools and an audit trail |
- Discover & inventory every AI tool, feature, model and agent in use — an AI register, like an asset register.
- Set a usage policy — which data classes may go to which tools, and who approves new ones.
- Provide an approved alternative that is at least as good as the shadow tool. This step decides whether the others work.
- Educate — explain the “why”: retention, training on inputs, prompt injection, wrong answers.
- Monitor, audit and enforce continuously — new AI features ship weekly; an annual review is already out of date.
3 · What Governed AI Looks Like
The approved path is an architecture, not a policy document. Every request — from a person or from an agent — passes the same checkpoints, in this order:
| Checkpoint | What it does | Shadow-AI risk it closes |
|---|---|---|
| Identity provider | SSO, MFA, conditional access for people; managed identities for agents | Personal accounts; agents with no owner |
| API gateway / WAF | TLS, rate limiting, IP filtering, bot protection | Abuse and cost blow-outs |
| AI gateway | One front door to approved models: routing, model allow-list, caching, usage and cost metering | Unapproved models and APIs; no visibility |
| Guardrails & DLP | Detect prompt injection and jailbreaks; find and mask personal data and secrets; filter unsafe output | Data leakage; manipulated output |
| Authorisation | RBAC/ABAC per user and per agent; token validation; least privilege on every tool call | Over-privileged integrations and agents |
| RAG & vector store | Retrieval respects the caller’s permissions; embeddings encrypted at rest | AI that reveals documents the user could not open |
| Observability | Central logs and audit trails of prompts, tool calls and decisions, sent to the SIEM | Actions nobody can reconstruct |
| Human in the loop | Approval step before high-risk actions (payments, deletions, external messages) | Autonomous agents acting without approval |
▸ The zero-trust rules apply unchanged: verify every request, give every identity least privilege, assume breach. An AI agent is just another identity — one that works faster and never gets tired.
4 · The Frameworks to Anchor It
| Framework | What it gives you |
|---|---|
| ISO/IEC 42001:2023 | A certifiable AI management system — the AI equivalent of an ISO 27001 ISMS, including an AI inventory and impact assessments |
| NIST AI RMF 1.0 (2023) | Four functions — Govern, Map, Measure, Manage — for AI risk, free to use |
| OWASP Top 10 for LLM Applications | The technical risk list: prompt injection, sensitive-information disclosure, excessive agency and more |
| India’s DPDP Act 2023 | Pasting customers’ personal data into an external AI is processing; the organisation stays answerable as data fiduciary |
| EU AI Act (Regulation 2024/1689) | Risk-tiered obligations for providers and deployers of AI used in or sold into the EU |
For the attacks themselves, see AI Security: New Attack Surface; for how the models work underneath, How LLMs Work, Explained.
Bottom line
- Banning drives it underground — the only durable fix is an approved path that is easier than the shadow one.
- Eight places, eight traces — gateway, DLP, OAuth, procurement and repository logs already hold the evidence.
- Agents turn a leak into an access problem — give every agent an identity, an owner and scoped tools.
- One front door — an AI gateway with guardrails makes every model call visible, governed and billable.
- Inventory first — ISO 42001 and the NIST AI RMF both start with knowing what AI you run.
Sources & method. Hiding places and risks reflect common enterprise AI-governance practice; detection and control mappings are the author’s analysis. Framework scopes from ISO/IEC 42001:2023, NIST AI 100-1 (AI RMF 1.0), the OWASP Top 10 for LLM Applications, the Digital Personal Data Protection Act 2023 and Regulation (EU) 2024/1689. Product-neutral: no vendor is implied by any checkpoint. Original analysis; no third-party graphics or text reproduced.
← Back to Cyber security