Zscaler 2026 AI Security Report Highlights Enterprise Data Exposure and AI Guardrail Failures
Report discovered via Awesome Annual Security Reports.
Summary
Zscaler’s 2026 report analyzes 989.3 billion AI/ML transactions from 2025, finding rising data transfers and extensive blocked disclosures. Red-team testing across more than 25 enterprise environments found critical flaws in every AI system tested.
Key points
- ThreatLabz analyzed 989.3 billion AI/ML transactions in Zscaler’s cloud during January–December 2025, reporting 83% year-over-year growth across more than 3,400 applications.
- Data transferred to AI/ML applications reached 18,033 TB, up 93% year over year; approved mainstream tools represented a major data-exposure surface.
- Organizations blocked 39.2% of AI/ML transactions as part of privacy, data-protection, and usage-policy enforcement.
- ChatGPT accounted for approximately 410 million DLP policy violations in June–December 2025. These represented policy-blocked attempts to transfer sensitive information, not confirmed data breaches.
- Across more than 25 enterprise environments, Zscaler reported critical vulnerabilities in every AI system tested, with a median time to first critical failure of 16 minutes.
- Red-team findings included data leakage, prompt injection, multimodal jailbreaks, and retrieval-related trust failures; changes in input language, encoding, or format could bypass guardrails.
- The report recommends inventorying standalone and embedded AI, auditing connected-data permissions, applying least-privilege access and inline DLP, validating model provenance, and continuously testing AI systems.
Article Details
- Publisher
- Zscaler
- Report Period
- 2025-01-01 to 2025-12-31; DLP and country-level data: 2025-06-01 to 2025-12-31
- Scope
- Global enterprise AI/ML usage, data protection, AI-assisted threats, and adversarial testing of enterprise AI systems.
- Sample Size
- 989.3 billion AI/ML transactions across more than 3,400 applications; red teaming across 25+ enterprise environments with more than 222,000 adversarial attacks, approximately 199,000 of which completed without error.
- Key Statistics
- AI/ML transactions increased 83% in 2025 compared with 2024, reaching 989.3 billion; the report notes a corrected year-over-year calculation.
- Data transferred to AI/ML applications reached 18,033 TB in 2025, a 93% year-over-year increase.
- Organizations blocked 39.2% of total AI/ML transactions in 2025.
- ChatGPT accounted for 410,181,006 DLP policy violations during 2025-06-01 to 2025-12-31; these represent attempted transfers stopped by policy, not confirmed data leaks.
- Zscaler red teaming identified at least one critical vulnerability in 100% of tested AI systems; median time to first critical failure was 16 minutes.
- Recommendations
- Maintain a continuously updated inventory of sanctioned, shadow, and embedded AI applications and functionality.
- Inspect AI prompts and responses inline and apply AI-aware DLP to prevent sensitive data exposure.
- Apply least-privilege access to AI interactions and disable auto-enabled AI features until reviewed and configured.
- Validate model provenance, datasets, dependencies, and grounding sources; audit permissions on connected repositories.
- Continuously red-team AI systems for prompt injection, jailbreaks, data leakage, and other exploitable weaknesses.
MITRE ATT&CK
Malware
Vendors
Codeium- **OpenAI dominates as the top LLM vendor.** OpenAI accounted for the vast majority of LLM-driven enterprise transactions (3x more than Codeium), establishing it as the current de facto LLM.OpenAI- **OpenAI dominates as the top LLM vendor.** OpenAI accounted for the vast majority of LLM-driven enterprise transactions (3x more than Codeium), establishing it as the current de facto LLM.Perplexity- **Perplexity** took the third position by transaction volume last year (12 billion transactions). Beyond AI-powered search, it also operates proprietary LLMs that power its answer engine. Accordingly, enterprise usageZscaler- [How Zscaler Delivers Comprehensive AI Protection](#how-zscaler-delivers-comprehensive-ai-protection)
Products
ChatGPT- **ChatGPT accounts for the overwhelming majority of DLP violations.** Across all AI/ML applications analyzed, ChatGPT generated 410 million data loss prevention (DLP) policy violations, affirming enterprise risks tiedCodeium- **OpenAI dominates as the top LLM vendor.** OpenAI accounted for the vast majority of LLM-driven enterprise transactions (3x more than Codeium), establishing it as the current de facto LLM.DeepL- **DeepL** continued to see strong adoption in global organizations (3.3%), supporting multilingual communication across business-critical content.GitHub Copilot| GitHub Copilot | 11,348,420,722 |Grammarly- **Integrated productivity apps anchor enterprise AI usage.** Grammarly became the #1 application by transaction volume, reflecting reliance on AI that operates directly within communication and business processes.Microsoft 365 Copilot- **EchoLeak** is described as a zero-click prompt injection style vulnerability in Microsoft 365 Copilot that could enable data exfiltration via normal email ingestion patterns.Microsoft Copilot- **Microsoft Copilot** rounded out the top five (3%), driven by its deep integration into Microsoft 365 and its role in automating daily productivity tasks.Zscaler Zero Trust ExchangeThe Zscaler Zero Trust Exchange™ platform provides robust, inline security controls designed to address the unique challenges of enterprise AI adoption:
Countries
Industries
EducationTechnology & Communication and Education saw the highest year-over-year increases.Finance & Insurance- **Finance & Insurance and Manufacturing lead enterprise AI usage again.** For the third year in a row, these sectors represented the largest share of AI/ML traffic (23% and 20%, respectively) behind theirHealthcareinformation such as financial records, personally identifiable information (PII), source code, healthcare data, and other regulated content attempts to leave the organization through an AI application and isManufacturing- **Finance & Insurance and Manufacturing lead enterprise AI usage again.** For the third year in a row, these sectors represented the largest share of AI/ML traffic (23% and 20%, respectively) behind theirTechnology & CommunicationTechnology & Communication and Education saw the highest year-over-year increases.