AI Security Gaps Expose Millions of Secrets as Traditional Frameworks Fall Short In 2024 and 2025, a wave of AI-related breaches exposed critical vulnerabilities in security frameworks designed for traditional systems. High-profile incidents—including the compromise of the Ultralytics AI library in December 2024, malicious Nx packages leaking 2,349 credentials in August 2025, and ChatGPT vulnerabilities enabling unauthorized data extraction—highlighted a growing disconnect between established security controls and AI-specific threats. These attacks resulted in 23.77 million leaked secrets in 2024 alone, a 25% increase from the previous year. Notably, the affected organizations had robust security programs, passed audits, and met compliance standards under frameworks like NIST CSF, ISO 27001, and CIS Controls. Yet, these frameworks, developed for conventional IT environments, failed to address AI-driven attack vectors. ### Where Traditional Frameworks Fail 1. Prompt Injection – Unlike SQL or XSS attacks, prompt injection manipulates AI systems using valid natural language, bypassing input validation controls that scan for syntax patterns. 2. Model Poisoning – Attackers corrupt training data during authorized processes, evading integrity controls designed to detect unauthorized modifications. 3. AI Supply Chain Risks – Pre-trained models, datasets, and ML frameworks introduce threats that traditional supply chain security controls (e.g., SBOMs, vendor assessments) cannot mitigate. ### Real-World Impact - The Ultralytics breach involved malicious code injected into the build pipeline, slipping past dependency scans. - ChatGPT vulnerabilities allowed attackers to extract sensitive data through crafted prompts, despite strong network and access controls. - Malicious Nx packages exploited AI assistants to exfiltrate secrets, weaponizing legitimate functionality in ways existing controls did not anticipate. ### The Compliance vs. Security Gap While compliance remains essential, it no longer guarantees protection. IBM’s 2025 Data Breach Report found that AI-specific attacks take longer to detect due to a lack of established indicators of compromise. Meanwhile, Sysdig’s research revealed a 500% surge in cloud workloads running AI/ML packages in 2024, expanding the attack surface faster than defenses can adapt. ### The Path Forward Organizations must go beyond compliance by: - Implementing AI-specific controls (e.g., prompt validation, model integrity checks, adversarial robustness testing). - Updating incident response plans to address AI threats like prompt injection and model poisoning. - Conducting AI-specific risk assessments to identify blind spots in existing security programs. Regulatory pressure is increasing, with the EU AI Act (2025) imposing fines up to €35 million or 7% of global revenue for violations. Yet, waiting for frameworks to catch up is not an option—proactive measures are critical as AI adoption accelerates. The threat landscape has evolved; security strategies must evolve with it.