2025年人工智能安全报告

Executive Summary3 AI's growing role in cyber threats29Supercharged social engineering29AI-driven malware and ransomware across the attack chain30Agentic AI: the next frontier in autonomous AI—andattack vectors31Case study: How threat actors are exploiting interest in AI33 Key Findings4AI and ML Usage Trends6AI/ML transactions overview6Blocked AI/ML transactions12Data loss to AI/ML apps13AI usage by industry14Industry spotlights15ChatGPT usage trends19AI usage by country20EMEA insights21APAC insights22 The Evolving Scope of AI Regulations35 AI Threat Predictions for 2025-202637 Best Practices for Secure Enterprise AI Adoption395 steps to securely integrate GenAI tools40 How Zscaler Delivers Zero Trust + AI42 Under the hood: Zscaler’s AI security and data advantage42A comprehensive approach to AI security43Leveraging AI security across the attack chain46 Enterprise AI Risks and Real-World Threat Scenarios23 Core risks of enterprise AI adoption23DeepSeek and open-source AI: the risk of frontiermodels in your pocket255 prompts to deception: DeepSeek-generated phishing page27 Research Methodology48About ThreatLabz48About Zscaler48 ExecutiveSummary_ Analyzing 536.5 billion transactions captured across the Zscaler ZeroTrust Exchange™ from AI/ML tools between February and December2024, ThreatLabz discovered both surprising and unsurprising shifts inusage trends by enterprises worldwide. Another year in the still new “era of AI” has come and gone, marked bygame-changing advancements, rising adoption across industries, andhigh-profile challenges. Enterprises now see artificial intelligence (AI) and machine learning (ML)as essential for growth, driving efficiency, smarter decision-making, andfaster innovation. On the other hand, AI adoption brings serious securityrisks, from unsanctioned usage (“shadow AI”) to data exposure. Evenmore concerning, threat actors seem to have the upper hand as theyweaponize these same tools to amplify attacks. What once required skillnow takes minimal effort. What once took hours now takes seconds. ChatGPT drove the most AI/ML transactions, making up nearly half ofthe total volume. From an industry perspective, the Finance & Insuranceand Manufacturing verticals drove the most transactions as top adoptersof AI. However, increased adoption didn’t mean unfettered access: alarge percentage of AI/ML transactions were actively blocked. Beyond usage trends, ThreatLabz discovered real-world threat scenariosfrom AI-enhanced phishing to fake AI platforms. This report alsoexplores recent developments in areas that will undoubtedly influence AIin 2025 and beyond, including agentic AI, the emergence of DeepSeek,and the evolving regulatory landscape. This shift was on full display in 2024. GenAI became a cybercriminal’ssocial engineering machine. Today, phishing emails mimic trustedcolleagues with eerie accuracy. Deepfake technology turns voices andvideos into weapons of deception. As AI/ML capabilities evolve and the threats they enable grow, theimperative is clear, more sophisticated, strong security controls, zerotrust architecture, and AI-powered defenses are no longer optional—they’re essential. Keep reading for more insights and actionable strategiesto help your organization securely adopt AI while staying ahead ofAI-driven threats. In 2025, the power and perils of AI loom larger than ever. Threat actorswill continue to push the boundaries of AI’s malicious capabilities. Yet,AI isn’t just enabling attacks—it’s also now a critical line of defense,powering the fight against these attacks. The Zscaler ThreatLabz 2025 AI Security Report examines the manyfacets of AI in cybersecurity, from AI/ML adoption to AI-driven threatsand security capabilities. Key_Findings ThreatLabzanalyzed 536.5 billion AI and ML transactionsin the Zscaler cloud from February 2024–December 2024.The key findings that follow are based on data spanning varying time periods* for comparative analysis. AI/ML tool usage saw an exponential rise year-over-year,with 36x more transactions (+3,464.6%) from 800+AI/ML applications in the Zscaler cloud,highlighting theexplosive growth in enterprise interest and dependence onthese technologies. Enterprises blocked 59.9% of all AI/ML transactions,reflecting concerns around AI data security and thesteps companies are taking in shaping their approachesto AI governance. ChatGPT remains the top application by transaction volume,accounting for nearly half of all AI/ML transactions (45.2%)from known applications,despite ongoing debates over itssecurity implications. ChatGPT is also the most-blocked AI application amongknown applications,followed by Grammarly, MicrosoftCopilot, QuillBot, and Wordtune, reinforcing growinginterest and caution when it comes to AI-powered writingand productivity assistants in enterprise settings. Enterprises are sending significant volumes of datato AI tools, with a total of 3624 TBtransferred byAI/ML applications. The Finance & Insurance an