Andrius Buinovskis is Head of Product at NordLayer, a toggle-ready network security platform for business.getty​Organizations are quick to adopt AI tools but often underestimate the cybersecurity implications. Without proper guardrails in place, AI use can quickly go from boosting efficiency to exposing businesses to cyberattacks and data leaks.​According to a report from IBM, 97% of organizations that reported experiencing AI-related cyber breaches disclosed that they didn't have AI access controls in place. In the report, businesses divulged that 60% of the breaches led to compromised data, and 31% disrupted operations.​The data indicates a concerning trend: Organizations are adopting AI without making sure that adequate security measures are in place.​For most employees, the browser serves as the primary gateway to popular, readily available AI tools. The ease of access creates a low barrier to entry, meaning that basically anyone can use these solutions without training or an understanding of the underlying security implications.​How AI Use Can Become A Liability​The first key risk to consider is unintentional data leakage. Users can input sensitive information, such as confidential documents, credentials, financial figures, customer details or intellectual property, into public AI chatbots. This information can potentially become part of the AI model’s training data or accessible to third parties, resulting in unintentional data leaks, severe compliance breaches (violating the GDPR or HIPAA, for example) and reputational damage.​Another common threat is indirect prompt injection. Cybercriminals often hide malicious instructions on public websites to carry out malware distribution campaigns. If a user copies this text and inputs it into a browser-based AI tool, the AI tool could execute the hidden malicious commands, like exfiltrating data from the browser (including data from other open tabs or authenticated sessions), initiating unauthorized actions on behalf of the user, such as sending out phishing emails, or triggering a malware download to the employee’s device.​Employees might also unknowingly install malicious AI browser extensions. For the average user, it can be challenging to differentiate a legitimate AI browser extension from a malicious one because the key indicator usually lies in excessive permission requests. Through the false promise of increasing productivity, malicious AI browser extensions can steal sensitive conversations from AI chatbots, intercept prompts, log browser activity or inject malware into the user’s device, all while operating within the browser’s environment and often without the employee’s knowledge.​Setting Up Guardrails​AI tools are becoming essential accelerators for efficiency and productivity. The liability, however, stems from their blind promotion and adoption across organizations. Encouraging employees to incorporate AI tools into their daily workflows without setting up a secure environment to do so or properly educating them on the risks can backfire and result in a cybersecurity incident.​Before widespread AI adoption, companies must firmly secure three core elements: observability, data loss prevention and access management. To safeguard browser-based AI tools, this requires following these steps:​1. Establish shadow AI observability.Each instance of employees using unauthorized AI tools creates a security blind spot, expanding the company’s attack surface. Having information on what web-based AI solutions employees are using enables security teams to identify threats, evaluate risks and inspect whether the tool is malicious.2. Adopt data loss prevention tools.Controlling clipboard access helps ensure that employees won't be able to paste sensitive company data into an AI chatbot. Prohibiting file downloads and uploads on unauthorized websites will protect users from accidentally downloading malware onto their devices, downloading excessive client data for malicious intent or uploading sensitive data to unsafe websites.3. Deploy data exposure tools.Going beyond controlling what employees can copy and paste into AI platforms, data inspection evaluates whether the information the user is trying to submit is sensitive, redacting it and forbidding users from submitting names, credentials or Social Security numbers.4. Embrace zero trust—both at the network and browser level.The zero-trust approach is to never trust and always verify—this principle is crucial when managing high-risk tools like AI. To prevent users from accidentally leaking sensitive corporate data, the first step is to reduce the resources that employees can access. Security teams can implement network segmentation to ensure that users are only accessing resources necessary for their roles, reducing the overall attack surface and limiting the scope of potential data exposure. If sensitive company data is accessible through a browser (stored in web-based software-as-a-service applications), security teams can restrict employee access to this data using browser traffic segmentation.​