security-review-owasp-ai-agent-security

To review an AI agent system for prompt injection vulnerabilities, analyze how user input, retrieved content, and memory interact across trust boundaries to identify paths enabling policy bypass or unauthorized actions. The review requires evidence-based findings with severity and confidence ratings.

Context leakage in AI agents occurs when sensitive information from memory, retrieved content, or system prompts is exposed through model outputs. Detect it by tracing data flows across the agent workflow and reviewing memory, logging, and retrieval paths for unauthorized data exposure.

Check AI agent tools for unsafe execution by analyzing tool invocation paths for least privilege enforcement, output validation, and containment. Identify abuse paths that could enable secret exposure, unauthorized actions, or unsafe autonomous behavior, documenting findings with remediation guidance.

Yes, you can assess AI agent operational guardrails and human-approval workflows by reviewing deployment controls, logging configurations, and approval triggers. The review evaluates whether guardrails adequately prevent unsafe autonomous behavior and enforce policy across the agent system.

The best way to report AI agent security review findings is to produce evidence-based reports with clear severity and confidence ratings for each issue. Separate confirmed problems from review gaps, documenting affected areas, evidence, impact, and remediation guidance for every finding.

Yes, this AI agent security review covers code, configuration, architecture, prompts, tools, retrieval paths, memory, logging, deployment controls, and human-approval workflows. It applies topic-specific checks for least privilege, output validation, and containment across all these areas.