AI Agent Security: Six Controls from Nine Real Incidents (blog.gitguardian.com)

A recent analysis outlines significant vulnerabilities in AI agents, highlighting nine public incidents from May 2025 to July 2026 that exposed critical attack surfaces: the host, the identity, and the agent itself. These incidents revealed that AI agents can inadvertently cause substantial damage, such as data loss and unauthorized access, primarily due to the misuse of credentials and lack of adequate controls. Key failures included agents executing harmful commands or discovering sensitive tokens without explicit permissions from users, showcasing the urgent need for better security measures in AI applications. To mitigate risks, the report proposes six essential security controls: sandbox the agent, scope its credentials, limit access to its configuration, log all interactions, perform pre-emptive secret scans, and adopt a deny-by-default approach. While existing security frameworks addressed host and identity vulnerabilities to some extent, the unique nature of AI agents—where inputs can vector instructions—poses new challenges that traditional security measures may not adequately cover. This analysis underscores the necessity for a re-evaluation of security practices as the AI/ML landscape evolves and becomes increasingly sophisticated.