AI Security
Overview
Xygeni's AI Security product detects security risks in the AI components used by your applications — prompts, agents, skills, tools, guardrails, models, and vector stores — directly from source code, before the application is deployed. It extends the visibility, prioritization, and control that Xygeni already provides for traditional software supply chain risks to the specifics of AI workloads and emerging AI threat models.
Every finding is classified under the OWASP Top 10 for LLM Applications and GenAI (2025) — the taxonomy auditors already know — and, where applicable, tagged with the OWASP Top 10 for Agentic Applications (2026, ASI) and the OWASP Agentic Skills Top 10 (2026, AST) categories, plus red-team attack vectors (Prompt Injection, Indirect Injection, Jailbreaks, RAG Poisoning, Multi-Modal Injection, Data Exfiltration, and more) for filtering and hunting.
Protect the AI Attack Surface of Your Applications
LLM SDKs, agent frameworks (LangChain, LangGraph, CrewAI, AutoGen, Semantic Kernel, LlamaIndex), MCP servers, and coding-tool skills expose new attack surfaces — tool execution, prompt injection, data exfiltration — that traditional SAST and SCA do not cover. Xygeni's AI Security scanner is built to uncover:
Prompt injection surfaces: system prompts that concatenate unbounded user-controlled content without role separation, delimiters, or guardrails.
Unprotected agents: tool-invoking agents without input guardrails, and agents ingesting external content (web fetchers, RAG retrievers, email/chat parsers) with no sanitization — the classic indirect injection path.
Mutable prompt sources: prompts pulled from a registry (LangSmith, Langfuse, etc.) pinned to a floating label (
@latest, a branch) instead of an immutable version.Multimodal exposure: agents accepting image, audio, or video input with no guardrail covering those modalities.
Obfuscated jailbreak payloads: prompt bodies matching known obfuscation classes (base64/hex decoding, Unicode smuggling, typoglycemia).
Malicious or typosquatted skills: coding-tool skill manifests (
SKILL.md,AGENTS.md) whose natural-language body hides malicious directives, and skill names that impersonate well-known skills or integrations.
Deterministic detectors are complemented by an optional LLM semantic review stage that judges the intent of skill and agent instruction manifests, catching prose-only attacks that no signature-based scanner can see.
For more information regarding AI Security, refer to these sections:
Last updated