Cisco Integrated AI Security and Safety Framework Report

AI threats now span four major categories beyond traditional cybersecurity: The framework identifies content safety failures (harmful outputs), model integrity compromise (poisoning and tampering), runtime manipulations (prompt injection), and ecosystem risks (multi-agent collusion)—requiring organizations to defend AI systems across the entire lifecycle, not just at deployment.

Existing security frameworks cover only "slices" of AI risks: While MITRE ATLAS, NIST AI 100-2, and OWASP Top 10s provide valuable guidance, the Cisco framework addresses a critical gap by integrating these perspectives into a unified taxonomy that covers AI security and safety across different AI modalities, agents, and pipelines simultaneously.

The framework is designed for practical implementation: Rather than remaining theoretical, the AI Security Framework is built specifically to support threat identification, red-teaming exercises, and risk prioritization—enabling organizations to move from understanding AI risks to actively managing them.

AI risks extend beyond individual models to entire ecosystems: The framework recognizes that threats include orchestration abuse and multi-agent collusion, meaning organizations must consider security risks not just within their own AI systems but across interconnected AI agents and platforms that may interact with each other.