Updated November 2026 · By the Folio3 AI Editorial Team · Reviewed by Abdul Sami, Head of AI Development
Unapproved AI tools risks fall into four groups: data exposure, security, compliance and bad decisions. The 12 biggest are data leakage, model training on company data, IP loss, malicious extensions, stolen credentials and unmonitored agents. The rest are compliance violations, missing audit trails, data residency conflicts, hallucinations, bias and vendor breaches. Most start with one employee pasting work data into a tool nobody reviewed.
IBM's 2025 Cost of a Data Breach Report found that one in five organizations had a breach tied to shadow AI, and heavy shadow AI use added about $670,000 to breach costs. IBM's Suja Viswesan put it plainly: "a gap between AI adoption and oversight already exists." New to the topic? Start with what shadow AI is.
The 12 risks at a glance:
# | Risk | What it puts at stake |
1 | Sensitive data leakage | Customer data, contracts, code |
2 | Model training on company data | Proprietary content |
3 | IP and trade secret loss | Patents, designs, roadmaps |
4 | Malicious extensions | Full chat histories, sessions |
5 | Exposed credentials and keys | Cloud and production access |
6 | Unmonitored agents and connectors | Mailboxes, drives, CRM |
7 | Compliance violations | Fines and failed audits |
8 | Missing audit trails | Ability to prove compliance |
9 | Data residency conflicts | Contracts and regulated deals |
10 | Hallucinated outputs and code | Decisions, sanctions, vulnerabilities |
11 | Bias in hiring and lending | Discrimination claims |
12 | Vendor and supply chain exposure | Breaches outside your control |
Unapproved AI tools are AI apps, accounts and features employees use for work without review by security, IT or legal. The tool can be legitimate; what makes it unapproved is that nobody checked its data terms, access controls or logging.
Quick answer: what counts as unapproved AI
If security cannot see it, log it or switch it off, it is unapproved. That covers free chatbots, personal logins, AI extensions, embedded AI features and agents. Some teams call these unsanctioned AI tools.
Unapproved AI vs shadow IT
Shadow IT stores data where IT cannot see it. Unapproved AI also processes that data, keeps it and may train on it, and it hides inside browsers, personal accounts and approved apps.
Common types of unapproved AI
The usual suspects are consumer chatbots, personal accounts on approved vendors, AI browser extensions, AI features switched on inside approved SaaS, and coding assistants or agents tied to personal accounts.
Why employees use unapproved AI
Employees use unapproved AI because the approved path is slower, weaker or unclear.
If the sanctioned tool lacks a feature or runs an older model, people open the free one in another tab. The fix is a better approved tool, not a stricter rule.
Slow approval and procurement cycles
Weeks of review for a cheap tool push people to personal cards or free tiers. Every workaround removes that tool from your inventory and from your controls.
Unclear or unknown AI policies
Microsoft's 2024 Work Trend Index found only 39% of AI users had received AI training from their company. A policy nobody has seen cannot guide anyone.
No ticket, no license, no onboarding. That friction gap drives bring your own AI (BYOAI) more than any other factor, and zero friction beats a policy that adds steps.
How widespread is unapproved AI?
Unapproved AI use is the norm, not the exception.
Source | Finding |
Microsoft Work Trend Index, 2024 | 78% of AI users bring their own AI to work |
UpGuard survey, 2025 | Over 80% of workers, and nearly 90% of security professionals, use unapproved AI |
Cyberhaven, 2024 | 73.8% of workplace ChatGPT use ran through non-corporate accounts |
IBM, 2025 | Only 37% of organizations have policies to manage or detect shadow AI |
Gartner, 2025 | Over 40% of organizations will face shadow AI incidents by 2030 |
Not sure which AI tools your teams already use?
Tell us how your teams work today, which AI tools you know about and what data they handle, and we will help you map the gaps before they turn into an incident. Call +1 408 365-4638 or email contact@folio3.ai.
Talk to Our Team
The risks fall into four groups: data, security, compliance and decision quality.
Sensitive data leakage through prompts
Cyberhaven found 27.4% of data employees put into AI tools was sensitive, led by support data, source code and R&D. Once pasted, it sits under vendor terms legal never reviewed.
Model training on company data
OpenAI does not train on ChatGPT Business, Enterprise or API data by default, but consumer accounts can be used unless the user opts out. The account tier, not the tool, decides your exposure.
Intellectual property and trade secrets
Pasting secrets into public tools weakens trade secret protection. IBM found shadow AI incidents compromised intellectual property in 40% of cases, against a 33% average across all breaches.
Malicious extensions and plugins
In December 2025, researchers found extensions capturing AI chats from over 8 million users, and two fake AI sidebars with 900,000 installs doing the same. A store badge is not a security review.
Exposed credentials and API keys
Developers paste code with hardcoded keys into chatbots. Group-IB also found 101,134 infected devices with saved ChatGPT logins for sale, each exposing saved chat history.
Unmonitored AI agents and connectors
An agent linked to Gmail, Drive or a CRM through a personal OAuth grant keeps reading data indefinitely. Custom AI agent development inside your environment adds scoped permissions and logs.
Regulatory and compliance violations
GDPR, HIPAA and PCI DSS follow the data, not the tool. IBM found shadow AI incidents exposed customer personal data in 65% of cases, against a 53% average.
Missing audit trails and logs
Consumer accounts give you no logs to export. When a regulator or client asks what left your environment, you cannot prove anything and must assume the worst.
Cross-border data residency conflicts
Free tools process data wherever the vendor runs. That breaks in-region contract terms and laws such as Saudi Arabia's PDPL, which restricts transfers outside the Kingdom.
Hallucinated outputs in decisions
Lawyers in Mata v. Avianca were fined $5,000 for ChatGPT-invented citations, with the judge stressing their "gatekeeping role" over accuracy. Unreviewed AI code merged into production carries the same risk.
Bias in hiring and lending
Illinois HB 3773, in force since January 1, 2026, bars employers from AI that has a discriminatory effect. A recruiter ranking CVs in a personal chatbot is covered.
Vendor and supply chain exposure
Wiz found a DeepSeek database open to anyone in January 2025, exposing chat history and API keys. You inherit every vendor's mistakes, including vendors you never assessed.
Operational problems with unapproved AI
Not every unapproved AI tools problem is a breach.
Data lost when employees leave
Offboarding closes company accounts, not personal ones. Prompts, automations and any company data shared through personal AI accounts leave with the employee.
Small AI subscriptions scatter across expense reports, so nobody sees the total, compares tools or negotiates a company rate.
Inconsistent outputs across teams
Different models and prompts give customers different answers, product claims and disclaimers, which means rework and statements legal never approved.
Fragmented knowledge and workflows
A prompt that saves an hour stays in one personal account instead of becoming a shared template, so the gain never scales beyond one person.
Incident response starts blind
The first days go to finding out who used what. IBM found extensive security AI and automation cut breach lifecycles by 80 days, so visibility directly buys time.
What most guides on unapproved AI miss
- An approved vendor is not an approved account. A personal login on the same tool skips every control you negotiated.
- Deleting a chat is not deleting the data. Vendors keep copies under their own retention rules. Samsung's 2023 memo cited this exact worry: data sent to AI platforms sits on external servers and is hard to retrieve or delete.
- Awareness does not reduce use. Security professionals use unapproved AI more than average, so training alone will not fix it.
- OAuth grants outlive intent. An agent connected once keeps access until someone revokes it.
Control | Approved AI | Unapproved AI |
Data terms | Reviewed, with a processing agreement | Consumer terms |
Training on inputs | Off by default on business tiers | Often on by default |
Access | SSO and roles | Personal email and password |
Logging | Central audit trail | None you can access |
Sensitive data | Redacted before the model | Whatever gets pasted |
Offboarding | Access revoked | History stays with the person |
Give your teams an approved AI path they will actually use
AI Guardian gives every team one controlled way into the models they need, with PII redaction, model access by line of business, and usage and cost visibility for each manager. Call +1 408 365-4638 or email contact@folio3.ai.
Explore AI Guardian
Regulations affected by unapproved AI
Unapproved AI breaks laws you already follow. For the full picture, see AI governance and compliance readiness.
GDPR and EU AI Act
GDPR fines reach €20 million or 4% of turnover, and Italy fined OpenAI €15 million in 2024. The EU AI Act's high-risk obligations now apply from December 2027.
HIPAA, PCI DSS, SOC 2
Health data in a tool without a business associate agreement breaks HIPAA. Card data in a chatbot escapes PCI DSS controls, and AI outside your SOC 2 controls creates audit exceptions.
Saudi and UAE PDPL
Saudi PDPL fines reach SAR 5 million per violation, and the UAE's Federal Decree-Law No. 45 of 2021 also limits cross-border transfers. Check where every AI vendor stores data.
US state AI laws
Illinois HB 3773 and Texas TRAIGA took effect January 1, 2026, and Colorado replaced its AI Act with SB 26-189 from January 2027. Watch any AI use in hiring or lending.
Real-world unapproved AI incidents
Year | Incident |
2023 | Samsung engineers pasted source code into ChatGPT; Samsung banned generative AI on company devices |
2023 | Group-IB found 101,134 stolen ChatGPT logins for sale |
2023 | Mata v. Avianca lawyers fined for fake citations |
2025 | DeepSeek database exposed chat logs and keys |
2025 | Extensions with 8 million and 900,000 users captured AI chats |
How to detect unapproved AI
Layer your signals, because no single tool sees everything. Gartner's Arun Chandrasekaran advises CIOs to "conduct regular audits for shadow AI activity."
Warning signs of shadow AI
Small recurring software expenses, uninvited meeting bots, sudden style changes in documents and shared chat links are early, free signals that unapproved AI is already in use.
Network and DNS monitoring
DNS and proxy logs give a fast first inventory of AI traffic, but they miss personal devices and AI inside allowed domains. Check local employee-monitoring rules first.
Browser extension and OAuth audits
Review installed extensions and OAuth grants in your identity and workspace consoles. This finds the AI sidebars, meeting bots and agents that network logs miss.
CASB, DLP and endpoint signals
CASB identifies and coaches on AI apps, DLP stops regulated data before it reaches a prompt, and endpoint tools find desktop AI apps. Together they turn discovery into enforcement.
Survey employees, not just systems
An anonymous, no-penalty survey reveals tools no log can see and, more importantly, why approved options fall short.
Rate each tool on its data and terms.
Classify data by sensitivity
Use four tiers: public (any approved tool), internal (company accounts only), confidential (processing agreement plus redaction) and restricted (dedicated environment or no AI).
Score each tool 0 to 2 on training, residency, processing agreement, attestations, SSO and logging. A zero on training or processing terms is a stop for confidential data.
Approved, limited and prohibited tiers
Approved tools handle confidential data, limited tools handle public or internal data for named teams, and prohibited tools are blocked with a suggested alternative and a review date.
A free meeting note-taker scores 3 out of 12, with zeros on training, processing agreement and SSO. Decision: prohibited for customer calls, limited to internal meetings.
Bans move AI use out of sight. Samsung's 2023 ban covered company devices and internal networks, leaving personal phones outside it. Make the approved path the fastest one instead.
Expert insight
“Shadow AI is a signal, not just a threat. Every unapproved tool points to a task your approved stack does not handle yet. Fix those first.”
Abdul Sami, Head of AI Development
Seven steps, in order.
Step 1: Build an AI inventory
Combine surveys, DNS logs, CASB data, OAuth grants and expense reports into one list showing every AI tool, its users and the data tier it touches.
Step 2: Write a usable AI policy
Keep it to two plain pages with one named owner, so a new hire understands it in five minutes. See common ownership gaps in AI governance challenges.
Step 3: Offer secure approved alternatives
Match each top shadow tool with an approved option that is at least as good. That is the core of AI enablement: better tools, not just rules.
Create a fast lane for low-risk tools and pre-approve common categories, so the official route becomes faster than the workaround.
Step 5: Control data, not websites
Use DLP and PII redaction instead of endless domain blocks. Compare options in our roundup of AI guardrail platforms.
Step 6: Train employees on safe use
Use real prompts from your own workflows and explain why each rule exists. Staff who understand the reason follow the policy without being policed.
Step 7: Monitor, review and update
Track approved-tool share, unknown apps found, DLP blocks and approval time every month. Budget the controls with our AI guardrails implementation cost breakdown.
What to do after exposure
Contain and scope the incident
Record the tool, account, data and time. Then delete the chat, rotate any shared secrets and revoke the account's sessions and OAuth grants.
Request vendor data deletion
Send the vendor's privacy team a written deletion request the same day, with the account and timestamps, and keep the confirmation.
Assess breach notification duties
GDPR allows 72 hours to notify the authority when people's rights are at risk. HIPAA, PDPL and client contracts set their own clocks, so bring in legal early.
How Folio3 AI can help
AI Guardian is an AI gateway for teams at mid-size companies. It gives staff one approved way into the models they want, with PII redaction before prompts reach a model, model access by line of business, routing to your own models or infrastructure, and usage and cost visibility for every manager.
Bringing unapproved AI into the light
The organizations handling this well do not have the strictest bans. They know which tools are in use, control what data can leave, and give people approved tools faster than the shortcuts. Unapproved AI tools risks shrink when the approved path becomes the easy one.
Ready to replace shadow AI with something better?
A short conversation is enough to review your AI use cases, data rules and existing stack, and agree on the first controls worth putting in place. Call +1 408 365-4638 or email contact@folio3.ai.
Book a Consultation
FAQs
The main risks are data leakage, model training on your data, IP loss, malicious extensions, stolen credentials, compliance violations and vendor breaches. IBM found heavy shadow AI use added about $670,000 to breach costs.
What is the difference between unsanctioned and unapproved AI?
They mean the same thing: AI used for work without review by security, IT or legal. Both fall under shadow AI.
Rarely, because use moves to personal devices with no visibility. A faster approved alternative plus data controls works better.
Is ChatGPT safe to use at work?
It depends on the account. Consumer plans can train on chats unless users opt out, while Business, Enterprise and API data is excluded by default.
Combine DNS logs, extension and OAuth audits, CASB and DLP signals, and anonymous surveys. No single method catches everything.
Are AI features inside approved SaaS apps a risk?
Yes, because a new AI feature can change how an approved tool handles your data. Treat each feature activation as a new review.
What should employees do after sharing sensitive data?
Report it to security the same day and rotate any shared passwords or keys. Early reporting leaves time for vendor deletion and notification deadlines.
How much does shadow AI increase breach costs?
IBM's 2025 report found high shadow AI use added about $670,000 per breach. Those incidents also exposed customer personal data more often than average.