Key Takeaways:
- Shadow AI happens when employees use AI tools without IT approval, often connecting them directly to company data and systems.
- Common sources include unsanctioned chatbots, machine learning tools, marketing automation, and data visualisation platforms.
- Security teams can detect Shadow AI through browser extension audits, CASB tools, data scans, and monitoring for unusual AI-related traffic.
- Clear governance, defined access controls, and an approved AI tools list reduce the temptation for staff to go looking for unsanctioned alternatives.
- Penetration testing adds an extra layer of defence by regularly checking for new integrations and weak points before they become a risk.
Introduction: What Is Shadow AI?
Shadow AI has become an increasingly prevalent issue in cybersecurity in recent years, accelerated by employees’ widespread adoption of AI tools in the workplace. While these AI tools can introduce new operational possibilities for workplaces, they also risk leaving organisations vulnerable to a range of cyber risks, such as data leaks and regulatory non-compliance. 93% of IT leaders express concerns about data security risks from AI tools. This blog is for security professionals and CISOs looking to understand and defend against Shadow AI.
What’s the difference between Shadow AI and Shadow IT?
Shadow IT is nothing new to the cybersecurity industry: it refers to employees using unregulated hardware, software, or cloud services without IT approval or oversight. Common examples of this include employees downloading SaaS tools unlicensed by their organisation without seeking approval or sign-off, or launching a CRM, or creating a personal Dropbox account.
Shadow AI falls under that umbrella, but with a critical differentiation: while shadow IT involves unauthorised hardware, SaaS applications, or cloud storage, Shadow AI actively processes and learns from enterprise data in ways that create much more menacing insider threats at scale.
The problem goes further than individual tools, though: people often connect those AI apps to company data, cloud storage, and internal systems, creating flows of data that security teams can’t actually see.
How Shadow AI Emerges: AI Tools and Platforms
Shadow AI can emerge from many common entry points in your online environment, including:
- Browser extensions
- Personal accounts (eg, personal Google accounts, personal email accounts)
- Decentralised purchasing, when employees buy or sign up for tools independently, without going through IT or procurement for approval, is also a notable entry point. Decentralised purchasing bypasses procurement controls and makes it difficult for IT teams to track what’s being used and whether it’s secure.
What does Shadow AI look like in Practice?
AI-Powered Chatbots
Using generative AI chatbots (ChatGPT, Gemini, Claude) that haven’t been signed off or authorised by their workplace is a form of Shadow AI.
Without proper formal oversight or employee training on safe AI use from your organisation’s IT team, employees are at risk of exposing sensitive company data when inputting prompts.
This puts the data at risk and jeopardises your data security, whether intentional or not.
Machine Learning Models for Data Analysis
When team members use external machine learning models to analyse data, they risk exposing sensitive business information to platforms they have little control over. This is common with tools used for customer analysis, financial forecasting, and performance reporting.
The risks grow when corporate datasets are used to train third-party models, as that data may be retained or repurposed without your knowledge. Without IT approval and proper vetting, organisations can unknowingly hand over valuable data with no way to get it back.
Marketing Automation and AI Features
Marketing teams often utilise AI for campaign optimisation. While AI systems can support teams in automating repetitive tasks and streamlining processes, they can also introduce security risks if not properly implemented or approved.
Not only do third-party marketing AI tools endanger your security posture, but they can also lead to compliance gaps with critical cybersecurity frameworks like ISO 27001, SOC 2, and PCI DSS, particularly where data handling, access controls, and vendor risk management requirements are concerned.
Data Visualisation and AI Capabilities
Similarly, the use of AI data visualisation tools to process company data can introduce security vulnerabilities without proper IT review.
When visualisation tools operate outside IT governance, there’s no guarantee the underlying data is accurate, up-to-date, or correctly interpreted, meaning business decisions could be based on flawed outputs. Without a review process, errors can go undetected and unchallenged until the damage is done.
Detecting Shadow AI: Monitor AI Features and Data Flows
Addressing Shadow AI and increasing visibility of your employees’ AI usage requires a mixture of both regular auditing and potential tooling adoption to bring to light any unsanctioned AI application use. Security leaders should:
Audit browser extensions and plug-ins for AI connections: Review every browser extension and plug-in in use across the organisation to identify any that connect to AI services, so you can spot unsanctioned tools before they become a data exposure risk.
Deploy a Cloud Access Security Broker (CASB) to surface unsanctioned AI app usage: A CASB gives you visibility into which AI applications employees are accessing, including tools that haven’t been approved by IT or security, so Shadow AI usage doesn’t stay hidden.
Scan data at rest for AI-relevant content:
- Run DSPM (Data Security Posture Management) scans to find sensitive data that’s at risk of exposure through AI tools.
- Flag over-permissioned files that are likely to flow into AI tools, so you can lock down access before sensitive information gets pulled into an AI workflow.
Monitor user behaviour for anomalies:
- Implement behavioural analytics to detect unusual patterns of AI access, such as sudden spikes in usage or access from unexpected accounts.
- Evaluate outbound traffic to known AI endpoints (e.g. api.openai.com) and check for anomalous patterns that could indicate Shadow AI activity. . Correlate these endpoint events with other signs of external AI connections to build a clearer picture of any unusual activity.
Preventing and Managing Shadow AI: AI Governance and Access Controls
As with all cybersecurity, prevention and proactive management are more effective than reactive remediation. By setting clear governance for your team, security leaders build a collaborative defence against Shadow AI, define responsibilities, and establish a clear response plan if Shadow AI is detected. Here are our expert suggestions on how to comprehensively introduce governance and access controls to your existing cybersecurity strategy:
| Control | What It Means | Who Owns It | Priority |
|---|---|---|---|
| Establish an AI governance framework | Define clear policies for AI use across the organisation, with named owners across IT, legal, compliance, and business teams | CISO / Head of IT | High |
| Define access controls based on role and data sensitivity | Restrict which employees can access which AI tools based on their role and the sensitivity of the data they handle | IT / IAM Team | High |
| Publish an approved list of AI tools and sanctioned alternatives | Give employees a clear, accessible list of approved tools so they have no reason to seek out unsanctioned alternatives | IT / Security Team | Medium |
| Implement risk-adaptive controls to reduce friction for safe users | Apply stricter controls where risk is higher, while keeping low-risk workflows smooth to encourage compliant behaviour | Security Operations | Medium |
| Enforce prompt-level DLP to block sensitive inputs to public models | Use data loss prevention tools to detect and block sensitive data being entered into public-facing AI models like ChatGPT | IT / DLP Team | High |
How Should I Respond to Detected Shadow AI?
While proactivity is critical when handling Shadow AI threats, responding appropriately to detected Shadow AI is also very important. It’s essential to try to minimise the spread of unsanctioned usage, and a distinctly outlined response plan is the key.
When responding to detected Shadow AI, security teams should:
- Contain discovered AI integrations immediately
- Audit exposed data and determine the impact scope
- Remediate by revoking access and patching integrations
- Communicate findings and corrective actions to stakeholders
- Update governance and training based on incident learnings
Measuring Success and AI Capabilities Improvement
With AI so frequently developing in its capabilities and workplace usage, having oversight into the efficacy of your AI governance can be immensely helpful to security leaders in keeping up with the turbulence. To measure success, we suggest teams:
- Track the reduction in unsanctioned AI incidents month over month to see whether your governance measures are working and where gaps still exist.
- Measure how long it takes to detect Shadow AI activity, using this as a benchmark to improve your monitoring and response processes over time.
- Monitor the percentage of employees using approved AI platforms versus unsanctioned ones to spot teams that may need extra guidance or support.
- Review training completion rates for AI use policies to make sure your workforce understands the risks and knows what’s expected of them.
How can OnSecurity help defend against Shadow AI?
Penetration testing plays a key role in Shadow AI defence by finding the gaps that unsanctioned tools slip through. OnSecurity’s continuous testing model means new AI integrations, browser extensions, and API connections get checked regularly, not just once a year, so Shadow AI doesn’t have months to sit undetected.
Our testers assess how AI tools interact with your systems, flagging over-permissioned access and weak points before they become a data leak. Combined with clear reporting that maps findings straight to your governance framework, teams get the evidence they need to tighten controls, update policies, and show auditors real proof of AI oversight, not just good intentions.
Get an instant quote today and find out how we can support your organisation in safe AI adoption.
