AgentProbe – adversarial security testing for AI agents (134 attack patterns) (github.com)

AgentProbe has been launched as a pivotal tool for adversarial security testing of AI agents, capable of executing 134 different attack patterns such as prompt injection, data exfiltration, and permission escalation. This tool functions similarly to OWASP ZAP but is specifically designed for AI agents, allowing developers to run automated security testing within CI systems before deployment. With an alarming 80% of IT professionals reporting unauthorized actions by AI agents, the introduction of AgentProbe addresses a critical need for enhanced security measures in AI systems, especially given the increase in enterprise AI deployments projected by Gartner. The implications of AgentProbe extend beyond simple testing; it provides a comprehensive suite of tests designed to uncover a range of vulnerabilities, including those that can arise during interactions between multiple AI agents. Additionally, the tool generates detailed reports and integrates with GitHub to highlight vulnerabilities directly within a repository's security tab, facilitating streamlined security management. By employing rule-based detection methods, AgentProbe ensures deterministic and efficient results without the cost of LLM dependencies, making it an essential addition to the AI/ML security landscape.