Shadow AI: Risks and Organization Examples

Published on: September 23, 2025

.

14 min read

.

For German Version

The practice of employing unauthorized software or technologies at work, or "shadow IT," is not new. But as AI has been more widely used, a part of shadow IT has developed into something more sophisticated: shadow AI. Workers now innovate without waiting for clearance thanks to AI-powered technologies like ChatGPT, AutoML platforms, and open source models. Although this may seem like a productivity victory, there are significant hazards involved.

Organizations adopting AI-driven solutions are becoming increasingly concerned about shadow AI as it functions beyond the purview of IT control. Even while it seems harmless, shadow AI can jeopardize compliance, expose private and sensitive data, and uncover security flaws that aren't readily obvious. Businesses run the danger of data breaches, legal infractions, and unrestrained AI-driven decision-making if they lack understanding of how AI is being utilized.

What is and is not shadow AI? Learn more about shadow AI, its emergence in customer experience and other sectors, its hazards, and how to mitigate them by using our guide.

Additional information in this guide:

• What is Shadow AI?

• Why Is Shadow AI Becoming a Concern for Businesses?

• What are the risks of shadow AI in organizations?

• How Do You Define Shadow AI in the Context of Enterprise IT?

• What are examples of shadow AI in real-world scenarios?

• How does Shadow AI differ from official AI deployments?

• Is Shadow AI Always Dangerous, or Can It Be Useful?

• What Causes Employees to Create Shadow AI Systems?

• How Can Companies Detect and Prevent Shadow AI Usage?

• What Role Does Shadow AI Play in Data Security and Compliance?

• Can Shadow AI Lead to Shadow Analytics and Unauthorized Insights?

• How does Shadow AI impact API security and shadow APIs?

• What are the challenges in managing shadow AI within IT policies?

• How to Control Shadow AI Without Stifling Innovation?

What is Shadow AI?

The term "shadow AI" describes the unapproved or poorly controlled usage of AI tools, models, agents, frameworks, APIs, or platforms inside a company that does not adhere to formal governance standards. Employees may use these AI technologies with the best of reasons, hoping to increase output or find more effective solutions to issues, but the absence of supervision poses serious operational, security, and compliance hazards.

Since these unscreened AI components have the ability to process sensitive data, make automated decisions, and introduce vulnerabilities that are missed by traditional security scanning and testing procedures, Shadow AI is a major blind spot in an organization's security posture from an AppSec standpoint.

Workers in a variety of industries use shadow AI, including as

• Email writing content generators
• Tools for AI analytics to assist with reporting
• AI-powered HR tools to evaluate job candidates
• AI-powered image producers
• AI coding helpers that are deployed privately
• Tools for risk assessment to examine fraud or credit risks

Why is shadow AI becoming a concern for businesses?

Shadow AI has security ramifications that go far beyond those of ordinary software. Workers who enter company data into unapproved AI systems run the risk of unintentionally disclosing private information to outside parties with dubious data management procedures.

Businesses may be at significant risk from poorly managed AI systems. These are the main obstacles:

• Exposure to sensitive information: Shadow AI frequently eschews standard security procedures, leaving companies open to threats from unprotected data access to data breaches.

  For instance, let's say a salesperson uses an AI application to assist them in summarizing important points for a meeting by pasting a customer contract. Unbeknownst to them, they may have given servers outside the company's control access to exclusive terms, client information, and pricing structures. Unauthorized parties may access this data or it may be included in the AI's training data.

  Although the behavior seems harmless enough, one of the biggest concerns connected to shadow AI is this kind of unintentional data leak.

• Compliance issues: Employees run the risk of breaking rules and nondisclosure agreements (NDAs) if they give sensitive information to outside AI platforms without the company's knowledge or consent.

• Absence of Monitoring: Critical failures that go undetected until harm is done might result from unvalidated models in production that are not retrained or monitored. As a result, companies may set up access restrictions to limit unauthorized applications and use network monitoring tools to keep tabs on program usage. Finding out whether and how unauthorized applications are being utilized may also be aided by routine audits and active communication channel monitoring.

  Before they become unregulated shadow AI threats, AI capabilities are appropriately evaluated and matched with business demands through regular monitoring and complete visibility. Because managing shadow AI necessitates enterprise-wide knowledge, proactive monitoring, and unambiguous governance.

  IT staff may keep an eye on the organization's adoption of AI by using SaaS management solutions that identify AI-powered apps.

• Operational Inefficiencies: When AI solutions are not properly integrated, they may result in data silos, technological debt, and incompatibility with current systems, rendering them unsustainable as requirements change.

• Data management problems: Accuracy, integration, and governance may be jeopardized by fragmented data from unapproved AI technologies, which might result in subpar insights and poor business choices.

• Risks to System Reliability: Shadow AI deployments have the potential to compromise system performance, maintenance, and dependability, making them crucial points of failure for operations and DevOps teams.

What are the risks of shadow AI in organizations?

While Shadow AI draws attention to the dangers of using AI without authorization, approved AI systems also provide security vulnerabilities, such as prompt injections and poorly designed plugins.

Six Major Shadow AI Risks.

1. Vulnerabilities in Security: When used outside of authorized contexts, shadow AI technologies circumvent company security measures and frequently don't integrate with endpoint protection, identification, and logging systems. Tools may overlook security patches or upgrades in the absence of official onboarding, raising the possibility of breaches, illegal access, and lateral movement. Additionally, certain technologies could bring model-level dangers, such as adversarial manipulation or prompt injection. The attack surface is further increased by the ability of AI plugins and extensions to seek enhanced permissions without inspection.

2. Data Privacy Violations: Employees who utilize unapproved AI technologies may unintentionally provide external systems access to regulated data, intellectual property, or sensitive company information. It is nearly impossible to trace, manage, or safeguard this data once it leaves the organization's regulated environment.

3. Regulatory Non-Compliance: By functioning without required risk assessments, transparency, or human oversight, the deployment of unapproved AI systems can violate laws such as GDPR, NIS2, and the EU AI Act. Failures to manage third-party tools, record AI use, or ensure the safe deployment of high-risk systems are all examples of this. It's possible to miss licensing limitations for third-party models or APIs, which might expose you to legal risks.

4. Exposure of Internal Data & IP: When staff members insert confidential information, such as papers, code, or designs, into unapproved tools, it may be saved or subsequently made public. Sensitive or classified information may inadvertently be revealed by even created outputs, particularly when tools are utilized improperly.

5. Governance and Oversight Gaps: When AI adoption is not managed across teams, there are oversight gaps since there is no centralized tracking, validation, or accountability. This restricts the IT, security, and compliance teams' capacity to identify AI use, evaluate risk, keep track of models and datasets, and guarantee audit preparedness. Debugging, legal defense, and confidence in AI-driven judgments are all made more difficult by the lack of a centralized AI model inventory and the inexplicability of output generation. Automated AI judgments without human-in-the-loop validation go against transparency standards and undermine accountability in high-stakes or regulated settings. These governance flaws expose the company to operational and regulatory repercussions by undermining controls mandated by frameworks like the NIST AI Risk Management Framework and laws like the EU AI Act.

6. Risks of the Supply Chain and Integration: Model dependencies, unreviewed extensions, integrated third-party APIs, and unauthorized connections to external AI services can all circumvent security evaluations and create supply chain risks. These include shadow integrations that erode control over AI behavior and outputs, the use of skewed or contaminated training data, and unsafe or unconfirmed model dependencies. Without adequate assessment, teams can potentially incorporate AI systems with opaque access scopes or ambiguous data handling procedures, creating blind spots in terms of security, compliance, or law.

How do you define shadow AI in the context of enterprise IT?

The term "shadow AI" describes the unapproved usage of artificial intelligence capabilities inside a company, frequently eluding security and IT monitoring procedures. Shadow AI appears when teams implement AI models, chatbots, or automation tools without the IT or compliance teams' knowledge, just like shadow IT occurs when staff members install unauthorized software.

The dangers of shadow IT are increased by shadow AI, which introduces predictive analytics, generative AI, and self-learning models that can function independently of IT. Workers may handle corporate data using external AI APIs, automation tools, or AI-powered assistants, frequently without knowing how these models keep, distribute, or utilize data.

Although both shadow AI and shadow IT use technology that is not directly under IT's control, they have different dangers and effects. While shadow AI particularly refers to artificial intelligence tools and models that operate without oversight, shadow IT refers to unapproved software or programs utilized by staff members.

Shadow AI raises additional issues with data privacy, automated decision-making, and threats associated with artificial intelligence, in contrast to shadow IT, which is mostly about the use of unapproved software. Without IT oversight, AI technologies may process sensitive corporate data, make predictions, and even automate activities, raising the risk of financial loss, security breaches, and compliance problems.

In contrast to shadow IT, which is frequently reserved for developers or tech-savvy users, shadow AI is used by staff members in all positions, the majority of whom are not knowledgeable enough to adhere to security best practices. A far larger and less predictable assault surface results from this.

Beyond conventional shadow IT solutions, a targeted strategy is needed to address shadow AI. Organizations must create governance that is suited to the particular dangers associated with AI, educate users, and promote teamwork.

Data is stored by standard software, but AI models may access it in novel ways. Sensitive data may be included in the model's training set, bringing it widely from the company's internal domain to the public realm. AI may produce choices that affect operations and potentially pose compliance problems as it handles sensitive corporate data.

Without adequate oversight, businesses run the risk of unintentionally disclosing confidential or client data, producing biased AI results, or breaking industry rules. In order to reduce risks and preserve innovation, monitoring shadow AI becomes an essential component of IT governance as AI capabilities continue to grow.

Moreover, alongside AI agents, identity and access management has two new challenges: establishing sensible security guidelines for erratic nonhuman actors and preventing an expanding army of malevolent agents from infiltrating corporate networks.

Large language models (LLMs) support AI agents, which are software entities that can employ tools to carry out multistep processes on their own. Although it is still in its early stages, as standard orchestration frameworks and agent-building tools advance, agentic AI is generally seen as the generative AI applications of the near future.

Supporting desired connections between AI agents and their tools, even those outside of a firm, without requiring IT personnel to set up authentication and permission for services beforehand, is another challenge for identity and access management in agentic AI settings.

What are examples of shadow AI in real-world scenarios?

Organizations use shadow AI in a variety of ways, frequently as a result of the demand for creativity and efficiency. Shadow AI can appear in a variety of ways depending on the organizational setting.

• Tools for generative AI: Using generative AI platforms such as ChatGPT, Claude, or Gemini is one of the most popular applications of Shadow AI. Without corporate supervision, staff members may use these technologies to write emails, produce content, write programming, or analyze data. Even while these technologies have amazing potential, processing sensitive corporate data with them has serious dangers.

  There may be pressure on a marketing intern to draft a press release as soon as possible. They replicate content that contains private client information, drawing inspiration from ChatGPT. The platform's data policy permits ChatGPT to keep user inputs for model enhancements, even though it produces an excellent draft. As a result, confidential client data is currently kept on external servers without the company's awareness.

  Technically speaking, developers may expose internal API keys in auto-suggested code snippets if they integrate GitHub Copilot into secure repository pipelines without secret screening. In a similar vein, developers who write documentation using OpenAI APIs may unintentionally reveal internal project code names or roadmap items in the information they produce.

• AI-driven code creation and evaluation: There are several concerns associated with developers using AI to create code snippets, SQL queries, or application logic from natural language descriptions.

  Security flaws, unsafe patterns, or implementation mistakes made by engineers without adequate scrutiny might be present in AI-generated code. Because AI code generation is so convenient, developers may implement solutions without fully considering the security consequences or doing sufficient security evaluations.

• Tools for marketing automation: Marketing teams may use shadow AI systems, which may analyze social media interaction data or automate email marketing efforts, to optimize campaigns. Using these strategies can result in better marketing results. However, in the event that client data is handled improperly, the lack of oversight may lead to violations of data protection regulations.

• Predictive modeling with machine learning techniques: Machine learning models may be used by data scientists and analysts to evaluate business data and produce forecasts. Without appropriate validation procedures, these models may introduce biases and mistakes that go unnoticed, access sensitive data, or produce untested outputs that affect business choices.

  Without realizing how using an external AI platform might lead to biased suggestions that alienate specific consumer demographics, a data scientist anxious to demonstrate the usefulness of predictive analytics for the sales department may do so. For instance, an analyst may unintentionally reveal confidential information when using a predictive behavior model to better analyze client behavior from a private dataset.

  Without realizing how using an external AI platform might lead to biased suggestions that alienate specific consumer demographics, a data scientist anxious to demonstrate the usefulness of predictive analytics for the sales department may do so. For instance, an analyst may unintentionally reveal confidential information when using a predictive behavior model to better analyze client behavior from a private dataset.

• Chatbots driven by AI: Teams in customer success may use unapproved AI chatbots to produce responses to questions. For example, instead of consulting their company's approved resources, a customer support agent may attempt to respond to a client's query by requesting responses from a chatbot. Inconsistent or inaccurate messages, possible consumer misunderstandings, and security issues if the representative's query includes private corporate information can all arise from this.

• Extensions for AI browsers: Workers may install browser extensions driven by AI that claim to increase productivity, automate activities, or summarize material. These extensions frequently grant wide access to browser data, which might lead to security flaws or the exposure of private data.

• Integrating LLM processes into apps: Without doing a thorough security analysis, developers frequently include big language models straight into apps. Retrieval-Augmented Generation (RAG) is a popular implementation paradigm that adds contextual information to AI replies by combining huge language models, embedding models, and vector databases.

• Deployment of local models: Open-source models are available for download and integration by developers without a security evaluation from repositories such as Hugging Face. Using machine learning frameworks, these models are frequently integrated straight into apps, enabling developers to carry out AI inference locally without relying on external APIs.

• AI bots that operate independently in DevOps processes: Autonomous AI agents that communicate with deployment, monitoring, or infrastructure systems may be deployed by DevOps teams. Without human assistance, these bots are able to assess system performance, identify irregularities, and even carry out corrective measures.

• Tools for data visualization: To swiftly build heat maps, line charts, bar graphs, and other data visualization tools, many businesses adopt AI-powered solutions. By clearly illustrating intricate data linkages and insights, these tools may support business intelligence. However, sharing business data without IT's consent may result in inaccurate reporting and possible data security and compliance problems.

How does Shadow AI differ from official AI deployments?

The use of artificial intelligence tools or systems without the consent, oversight, or participation of an organization's security or IT departments is known as "shadow AI."

Employees frequently utilize AI applications for work-related activities without their employers' knowledge, which can result in security hazards, data breaches, and compliance problems.

IT staff oversee security, compliance, and integration with current systems in the organized implementation stages of traditional AI deployment. Conversely, shadow AI infiltrates the company unchecked, frequently via individual workers or business divisions.

Organizations need to find a balance in order to securely harness AI's commercial potential. While taking use of its revolutionary potential, promoting responsible adoption inside safe frameworks can stop the development of shadow AI.

Is Shadow AI always dangerous, or can it be useful?

Yes, it is dangerous. Most uses of shadow AI start off with good intentions. Employees are trying to save time or become more productive. Companies are embracing digital transformation, which involves rethinking processes and decision-making by using AI technologies. Employees should utilize AI under the guidance and expertise of their IT security team for these reasons.

Organizations may take into account a number of strategies that promote responsible AI use while acknowledging the need for adaptability and creativity in order to mitigate the hazards associated with shadow AI. For instance, a deeper understanding of AI's potential and constraints may be facilitated by open communication between business divisions, security teams, and IT departments. In addition to ensuring adherence to data protection procedures, a collaborative culture may assist firms in determining whether AI technologies are useful.

Second, restrictions on AI use can act as a safety net, assisting in making sure that staff members only utilize authorized instruments within predetermined bounds. Policies governing the usage of external AI, sandbox environments for testing AI applications, or firewalls to prevent access to unapproved external platforms are examples of guardrails.

Additionally, eliminating all instances of shadow AI may not be possible. As a result, companies may set up access restrictions to limit unauthorized applications and use network monitoring tools to keep tabs on program usage. Finding out whether and how illegal applications are being utilized may also be aided by routine audits and active communication channel monitoring.

Lastly, the field of shadow AI is always changing, posing fresh difficulties for businesses. Employers may educate staff members about shadow AI and the dangers it poses by implementing frequent communications, such newsletters or quarterly updates. It goes without saying that the IT security team should oversee and be aware of all of this.

In actuality, the issue is not motivation. This is because these actions circumvent the usual review and approval process. If shadow AI is not sufficiently regulated or minimized, it presents serious organizational risks. For example,

• Employees who provide sensitive data to external AI platforms without the company's knowledge or authorization risk violating regulations and non-disclosure agreements (NDAs).

• Due to shadow AI's lack of oversight and validated security requirements, corporate data may be compromised or at danger of manipulation, and unapproved and unvetted tools may introduce errors, viruses, or flawed code into company processes.

• Isolated and incompatible AI systems may produce inconsistent or unreliable results, which might harm the reputations of customers or employees.

• Organizations may promote a culture of ethical AI usage by raising awareness of the consequences of employing unapproved AI tools. This knowledge might motivate staff members to look for authorized substitutes or speak with IT before implementing new software.

What causes employees to create Shadow AI systems?

Despite the concerns, workers in a variety of industries utilize shadow AI, from retail to customer success and beyond, since it:

• Speed Over Process: Teams place more emphasis on quick fixes or creative solutions than on official approval procedures.

• Unmet Internal Needs: When internal choices are delayed, unreliable, or feature-poor, employees resort to external AI technologies.

• AI Tools' Accessibility: More AI tools are now accessible than ever before, and many of them are free or inexpensive. Thanks to independent platforms or built-in capabilities in pre-existing software, employees may now easily incorporate AI models into their workflows. Additionally, AI technologies may be used through online interfaces without the need for infrastructure or installation.

• Lack of Awareness: Many workers are not properly trained to utilize AI, which might result in unintended hazards. Employees may trust AI-generated outputs without verification, upload sensitive data into AI models, or overlook security issues related to AI-generated content if they are not properly guided. One factor contributing to the unchecked proliferation of Shadow AI is a lack of AI literacy across departments.

• Lack of a Clear Policy: AI usage standards are either nonexistent, ambiguous, or inadequately explained.

• Experimentation Culture: Teams are urged to "try fast, fail fast," which results in the uncontrolled adoption of AI. Pressure on Marketing and Productivity: Business divisions use AI to increase productivity or stay ahead of the competition while avoiding supervision. Without waiting for IT to review and approve technologies, workers utilize AI to create content, automate repetitive processes, and analyze big information. Shadow AI deployments that eschew risk evaluations may arise from this speed-first mentality, raising the possibility of data security flaws and legal infractions.

• SaaS Integration: AI functions in programs like Slack, Notion, or Canva are turned on without consent or approval.

• Limited Governance Coverage: AI products infiltrate enterprises without security evaluations, compliance checks, or purchase permissions when there are unclear governance mechanisms in place. The lack of clear regulations for the use of AI in many organizations leads to disparities in the adoption and management of AI solutions by various teams. AI capabilities may go unnoticed or unassessed in security and procurement evaluations, which might result in unchecked adoption across a variety of company tasks.

• The Diversion of Employee and Business Vision: AI is becoming more and more popular among businesses and employees. Employees could, however, implement AI technologies that cater to personal needs rather than corporate objectives. This may lead to disjointed and inconsistent deployment of AI.

How can companies detect and prevent Shadow AI usage?

Establishing AI governance guidelines, keeping an eye on AI adoption, educating staff members about AI hazards, and incorporating AI supervision into security and compliance frameworks are all recommended. Unauthorized AI usage can be found with the aid of SaaS management solutions and routine audits.

By proactively reviewing third-party terms of service and product specifications, as well as by carrying out continuous monitoring and routine audits, shadow AI can be identified:

• Upfront reviews: Examining vendor technology's terms of service in order to assess embedded AI

• Continuous observation: Keeping an eye on model usage logs for improper and unapproved uses

• AI audits: Regularly examining how employees in a company are using AI

Organizations must take proactive and continuous measures to lower their risk exposure.

A clear and effective AI governance policy supports the mix of technological and human solutions used to prevent shadow AI.

Adopt Compliance and AI Training Measures:

Organizations may choose to adopt compliance and AI training measures as part of a number of strategies to mitigate the dangers associated with shadow AI.

• Implement AI training for staff members: HR departments should collaborate with compliance and legal to teach staff members. Information regarding artificial intelligence (AI), its applications, its hazards, and the technological capabilities that the company has approved should all be included in this literacy course.

• Distribute an AI Acceptable Use Policy: The company should specify exactly what it considers to be appropriate AI use. Usually, this is outlined in an Acceptable AI Use Policy that lists approved technology, permitted applications of AI, a review and approval procedure for AI usage, and sanctions for infractions.

• Make personal device usage apparent: To lessen the possibility that shadow AI will be used to finish tasks on unmonitored devices, organizations should make it plain that exporting company data or carrying out business operations on personal devices is not permitted.

Invest in technological solutions such as AI discovery:

Shadow AI is a shifting target that requires more than just a checklist to manage. Consequently, make investments in technical solutions like AI discovery as outlined below.

• Limit access to unapproved tools: IT teams should be careful to restrict and deny access to AI solutions that the company has not approved. This might entail restricting employee access to particular websites or portals or purposefully disabling suppliers' AI capabilities inside their current technological stack. Limiting access to certain AI technologies to personnel who have received enough training might be another example.

• Block sensitive data transfer, such as personally identifiable information: IT departments can use software that proactively prevents sensitive data from being transmitted before it is included in the AI solution for approved AI tools. This can significantly lower the chance that PII will be disclosed to third parties, even while it cannot ensure that sensitive information will be protected.

Put in place a program for AI governance:

Managing shadows effectively AI calls for more than simply financial investments in technology. It entails developing a plan for the administration of AI, streamlining and tracking its use, and integrating several tactics. You should adhere to these strategic best practices.

• Create a plan for AI governance: AI governance initiatives may assist companies in monitoring and controlling their authorized AI in both internal and external technology. They lay the foundation for regular evaluation of AI applications and guarantee regular and continuous evaluation as the field's capabilities and landscape evolve quickly.

• Simplify and monitor AI use: IT departments may improve visibility and keep tabs on their firms' AI by using an AI governance platform such as FairNow. Platforms for AI governance lower adoption risks and promote transparency.

An organization cannot completely eliminate the risk of shadow AI by using a single strategy, but by combining many strategies, possible effects can be significantly decreased.

What role does Shadow AI play in data security and compliance?

Shadow AI poses significant cybersecurity risksas workers could inadvertently shareprivate information into AI models that store or handle data from outside sources. AI-generated outputs might potentially be employed in important business decisions without enough validation if IT controls are not in place.

Workers may enter private, regulated, or proprietary data into external AI systems without understanding where or how that data is kept. It happens as a result of certain technologies keeping metadata or inputs on external servers. Data may be exposed without anyone's knowledge if an employee utilizes them to execute internal code or client information.

However, when data handling standards are ambiguous, AI models that handle and keep company data may be in violation of industry rules like GDPR, HIPAA, and SOC 2. This can lead to penalties, inquiries, or legal action. Because businesses find it difficult to monitor where data is being processed, stored, or utilized in AI processes, shadow AI may lead to inadvertent [compliance](/docs/network-security-tutorials/what-is-cybersecurity-compliance) violations.

Businesses in highly regulated sectors are particularly vulnerable because unapproved AI models might provide inaccurate financial reports, biased hiring choices, or unapproved medical advice, which could result in fines and legal ramifications.

How does Shadow AI impact API security and shadow APIs?

Shadow AI technologies are able to get beyond data management regulations set out by the DPDP Act, GDPR, and HIPAA. This can lead to penalties, inquiries, or legal action. These technologies frequently introduce uncontrolled integrations, personal device access, and unsafe APIs. Attackers may use any of those as a point of entry.

API keys, which are usually needed for AI services, could not be adequately protected in code repositories or configuration files. Security flaws might arise if developers using Shadow AI unintentionally reveal API keys in configuration files, code, or logs.

Unauthorized use, possible data breaches, and unforeseen service fees are all consequences of exposed API keys. The danger of exposure is increased because typical security controls for secret management may not be enforced since these implementations take place outside of governance frameworks.

While writing a report, an employee may copy and paste private information into a chatbot. Without informing IT, a team may develop internal tools using an open-source LLM API. Using services like Hugging Face or OpenRouter, developers may include GenAI capabilities into pipelines or apps. Others may connect to SaaS programs with built-in AI capabilities using personal accounts.

What are the challenges in managing shadow AI within IT policies?

Shadow AI creates additional security, compliance, and governance issues that enterprises need to handle, just like shadow IT brought uncontrolled software vulnerabilities. Without insight into how AI is being used, businesses risk data leaks, legal violations, and unchecked AI-driven decision-making. Because of this, handling shadow AI poses a number of significant difficulties, all of which have the potential to impact operational effectiveness, security, and compliance. A data table outlining these major issues and their implications for AI governance and corporate operations can be seen below.

• Absence of Visibility: BIT and security personnel may be put under stress when resources are diverted to handle or fix problems brought on by unlawful AI use. Because shadow AI frequently functions without IT supervision, it can be challenging to monitor AI systems, track AI usage, and manage risks.

• Data Privacy Issues: Unauthorized AI applications may result in the improper treatment of private information, putting the business at risk of data breaches or penalties from the government.

• Unauthorized AI Models: Employees could use untested AI models, which raises the possibility of biased judgment and erroneous findings.

• Vulnerabilities in Security: Shadow AI raises the possibility of cyberattacks by introducing possible security vulnerabilities like unprotected access points.

• Absence of Responsibility: It is difficult to determine who is accountable for mistakes, data breaches, or system malfunctions brought on by shadow AI in the absence of official monitoring.

How to control Shadow AI without stifling innovation?

Organizations must adopt a proactive strategy that includes leadership support, policy creation, staff training, and ongoing monitoring in order to effectively manage shadow AI. AI technologies have the potential to cause security threats, noncompliance with regulations, and unforeseen financial strains if they are not properly governed. The following tactics assist companies in managing shadow AI while preserving creativity and efficiency.

• Teaching Leadership and the C-Suite about the Dangers of Shadow AI: In the governance of AI, executive leadership is essential. However, a lot of businesses have trouble explaining the whole range of dangers that come with shadow AI. To make sure they comprehend the potential effects of uncontrolled AI deployment on data security, compliance, and operational efficiency, IT teams must interact with department heads and C-suite executives.

  Shadow AI governance necessitates cross-functional cooperation from the legal, financial, compliance, and human resources departments in addition to IT participation. A more organized, strategic approach to AI supervision may be developed by enterprises by making sure that leadership is in agreement on the risks and duties related to AI.

• Creating Regulations and Guidelines for Generative AI Tools: The use of AI is growing, and IT executives understand that formal control is necessary. However, constant enforcement, frequent updates, and staff understanding are necessary for these rules to be effective.

  Clear policies defining the usage of AI tools, which technologies are authorized, and data handling procedures should be established by organizations. To make sure AI deployment complies with security best practices, governance models should incorporate risk assessments, procurement approvals, and compliance checks.

• Employee Awareness and Communication Regarding Shadow AI: Many workers utilize AI-powered technologies without considering the hazards or if doing so is permitted by corporate policy. IT departments need to create effective communication plans to make sure staff members comprehend:

  - Which artificial intelligence techniques are permitted and which are not

  - How to check the correctness of AI-generated material

  - What security threats are present while utilizing automation techniques driven by AI?

  - How to submit the use of shadow AI for appropriate review

  - Organizations can lessen the possibility of inadvertent security lapses or compliance issues linked to unapproved AI products by informing staff members.

• Track the Use of AI: Fortunately, there are methods available to identify artificial intelligence in apps. Other AI models trained for that purpose may find specific features in the files and code of AI models and agents. Similarly, these models are able to identify open-source AI model licensing.

  Mend AI, for example, looks for hidden AI components in codebases, application manifests, and dependency trees. Following that, it produces an awareness report (Shadow AI report) that offers a comprehensive organizational map of AI usage, giving insight into the extent of AI use across various projects, products, and organizational divisions.

  An essential first step is to put in place monitoring systems that can identify AI-related activity across networks, apps, and cloud services. These monitoring tools ought to be able to recognize:

  - API connections to external services pertaining to AI

  - Applications using machine learning frameworks and libraries

  - AI components and model files in container images

  - Transferring data to AI platforms and services

  - Embedding services and vector databases

• AI Tool Inventory and Audit: A baseline for governance initiatives is established by carrying out an exhaustive audit to determine all AI models and technologies utilized within the company. Information on which AI systems are being utilized, by whom, for what reasons, and what data they analyze should all be included in this inventory. AI artifact discovery for model files, configuration files, training datasets, and LLM fine-tuning checkpoints should also be a part of this.

  After that audit, keep an internal AI Asset Registry or source-of-truth for each AI deployment and model.

  You may identify a variety of AI technologies with a tool like Mend AI, including as embedding libraries, open ML models from registries like Hugging Face & Kaggle, and third-party LLM APIs like OpenAI and Azure. This gives you complete insight into the AI elements—including Shadow AI—that are utilized in your code, allowing you to spot and report any instances of usage that are not authorized by the registry. More efficient risk assessment and mitigation techniques are made possible by this inventory, which offers vital insight into the organization's AI attack surface.

• Technical AI Governance Implementation: AppSec teams must put in place thorough governance controls:

  - Integrating AI security checks into CI/CD pipelines to identify and assess AI components throughout the development and deployment process is known as CI/CD pipeline integration. Prior to deployment, these checks can enforce governance norms, verify security settings, and detect unapproved AI components.

  - Using AI-aware dependency constraints to limit which AI packages, libraries, and models may be used in applications is known as dependency governance. This covers version pinning, authorized repository setups, and automated vulnerability scanning for AI components.

  - Network Controls: To limit which external AI services apps may connect with AI API endpoints, egress filtering should be implemented. This comprises proxies, network policies, and API gateways that implement access controls for interactions with AI services.

• Putting Technical Guardrails in Place: As IBM points out, from a technological standpoint, guardrails can include firewalls to prevent unauthorized external platforms, sandbox environments for testing AI applications, and restrictions surrounding the usage of external AI.

  The following are technical measures for secure use of AI:

  - Using organizational proxies for AI services to arbitrate communication between apps and external AI services is known as proxy services for AI APIs. These proxies may log interactions, filter sensitive data, enforce security regulations, and offer centralized administration for the use of AI.

  - Container Security Policies: To enforce security rules for AI workloads, policy engines such as Open Policy Agent (OPA) are implemented. These regulations can impose security setups, limit the types of AI models that can be used, and guarantee adherence to company standards.

  - Providing authorized settings for AI development with pre-approved tools, libraries, and services is known as "secure AI development environments." By giving developers the tools they require and enforcing security rules, these environments might lessen the motivation for developers to use alternatives to Shadow AI.

• Put Access Controls in Place: Role-based access controls (RBAC) should be put in place for AI tools that handle security-sensitive activities, and input and output logs should be routinely audited to identify any data exposure.

  Shadow AI dangers can be considerably decreased by limiting access to private information and putting in place safeguards against illegal data exchange with outside AI services. Technically speaking, these controls might consist of:

  - Instruments for preventing data loss that identify and stop sensitive data transfers

  - Filtering network traffic for AI service endpoints

  - API gateways that make sure AI services have access controls

  - Policies for container security that limit AI workloads

  - Protect critical AI processing in secure enclaves. Keep an eye out for illegal usage of AI model hosting platforms (such as AWS SageMaker and Azure AI).

• Planning for Incident Response: Creating incident response procedures, especially for security issues involving AI, guarantees that the company will be able to react efficiently in the event that Shadow AI results in data exposures or other security breaches. These procedures must consist of:

  - Detection Mechanisms: Put in place monitoring for abnormalities unique to AI, such odd trends in API usage, dubious data transfers, or unexpected model behaviors.

  - Methods of Isolation: Defining containment techniques, network isolation, and service suspension as ways to isolate corrupted AI components. Remove access tokens, cut API keys, and take a snapshot of the impacted resources.

  - Eradication: Clear out any leftover artifacts from repositories, containers, and cloud storage, as well as any unapproved models, extensions, or services.