AI Agents Can Work Faster Than Humans and Fail Harder Too (www.forbes.com)

AI agents are leaving labs and entering production—automating code, finance, and ops at machine speed. That speed is the point and the peril: agents can execute thousands of actions per second, so mistakes or malicious prompts can cascade far faster than traditional human-centered controls can catch them. Current permission models—static roles, hard‑coded logic and spreadsheets—were built for slow human workflows and are ill‑suited to continuous, autonomous agents. The rush to prove ROI for generative AI compounds the risk, as teams often deploy agents with human-equivalent privileges that lack contextual judgment, creating single-key breaches, unintended escalations, and compliance exposure. The solution is technical and architectural: treat authorization as the deterministic, API-driven layer that contains probabilistic models. Implement automated least‑privilege (time‑bound, task‑scoped permissions), access-as-transaction semantics, human-in-the-loop approval and strong MFA for high‑risk actions, continuous monitoring, and comprehensive logging for auditability. Vendors like Oso Security are pushing modular authorization APIs to replace bespoke, scattered logic. For enterprises the implication is clear: governance and adaptive access controls—not bans—will enable safe autonomy. Organizations that redesign permission models and telemetry will gain speed with manageable blast radii; those that don’t risk costly failures, regulatory scrutiny, and stalled innovation.