This Leadership Compass is an update of last year's Generative AI Defense. To see what it entailed, please click here. This update analyzes solutions that protect organizations from the distinct attack surface introduced by generative AI: prompt injection, jailbreaking, model manipulation, unsafe or harmful output generation, and data exfiltration through AI interfaces. It also covers the emerging threats arising from agentic AI workflows, including multi-agent orchestration, AI tool-use chains, and Model Context Protocol (MCP) security.
The focus of this Leadership Compass is on solutions that provide dedicated, generally available security capabilities across one or more of the following functional areas:
- Prompt and guardrail security: input validation, injection prevention, jailbreak detection, output filtering, and content safety enforcement
- Model integrity and red teaming: adversarial testing, model behavior validation, vulnerability assessment, and continuous model security monitoring
- AI data security: training data protection, sensitive data detection, and data leakage prevention via AI interfaces
- AI Security Posture Management (AI-SPM): AI asset discovery, risk posture scoring, misconfiguration detection, and compliance reporting
- Agentic AI security: agent identity verification, tool-use monitoring, MCP security, and multi-agent orchestration controls
Main capabilities include but are not limited to: prompt injection detection and prevention; jailbreak and adversarial input detection; AI output safety and content filtering; data leakage prevention via AI interfaces; model integrity verification and tamper detection; AI Security Posture Management; agentic AI threat detection and control; multi-model and multi-provider coverage; real-time runtime protection; and integration with SIEM, SOAR, and broader security platforms.
Inclusion Criteria
Solutions that provide dedicated, generally available capabilities for detecting, preventing, or mitigating threats that specifically target or exploit generative AI systems, including LLMs, multimodal AI, and AI agents. Solutions must address at least one core functional area: prompt injection or jailbreak detection; AI output safety enforcement; model integrity verification; AI data leakage prevention; AI Security Posture Management; or agentic AI threat detection and control.
Solutions must be actively marketed and sold as AI security products, with documented threat models specific to generative AI, not general-purpose security tools with incidental AI coverage.
Exclusion Criteria
Vendors that only cover traditional application security, generic web application firewalls (without AI-specific threat models), standard data security posture management (without documented AI asset coverage), or general content moderation without an adversarial security framing are excluded.
AI governance and responsible AI tools focused solely on fairness, bias, and explainability, without a security threat detection or prevention component, are out of scope.
However, there are no further exclusion criteria such as revenue or number of customers. We cover vendors from all regions, from start-ups to large companies.
