AI Security Gaps: Third-Party Agents Evade Control

Many third-party products now embed AI, but over a thousand operate outside identity infrastructure, posing a security challenge. Traditional AI security controls rely on a decision point, but agents arrive embedded in existing software without explicit adoption. This gap is critical as agents can act autonomously within enterprise systems, inheriting permissions from platforms like Slack. Security leaders must consider three agent origins: inherited, configured, and built, with the first two being the most prevalent and risky. The real risk lies not in the AI models but in the scaffolding and ecosystem they operate within. Effective governance requires continuous monitoring and understanding of agent reach and actions.