Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

AI agent adoption and development are evolving quickly. The tooling used to build agents is improving fast, but the security controls around those agents are often rigid, opaque, or difficult to adapt to real environments. As more teams experiment with OpenClaw, one challenge becomes clear: developers need ways to inspect what agents are doing, evaluate risky behavior, and intervene when necessary.

AI agent risk doesn’t emerge in a single moment. It develops over time across configuration changes, runtime behavior, long-horizon tasks, and interactions between agents, users, and enterprise systems. Their behavior and exposure can shift in real time as agents rewrite instructions, update memory, and dynamically alter execution.

How the integration works: Zenity integrates with the Foundry control plane to inspect agent behavior and enforce security policies inline at runtime. Over the past year, Microsoft Foundry has emerged as a cornerstone for enterprises building and deploying homegrown agents at scale. Organizations across industries are using Foundry to move beyond experimentation and into production, creating AI agents that can reason, invoke tools, access enterprise data, and automate complex workflows.

One recurring theme in my research and writing on agentic AI security has been the distinction between soft guardrails and hard boundaries. As someone who serves on the Distinguished Review Board for the OWASP Agentic Top 10, and who spends every day thinking about how to secure agents across enterprise environments at Zenity, this distinction is not academic. It is potentially the single most important conceptual framework practitioners need to internalize right now.

AI agents are rapidly becoming embedded in enterprise workflows, influencing revenue operations, customer engagement, development, and internal decision-making. As these systems gain autonomy and inherit access across SaaS, cloud, and endpoint environments, they introduce a new layer of operational and security risk that traditional controls cannot fully manage.

How a routine calendar invite enabled silent local file access and data exfiltration Note: This post is part of a coordinated disclosure by Zenity Labs detailing the PleaseFix vulnerability family affecting the Perplexity Comet Agentic Browser. This blog focuses on browser-level autonomous agent execution and session compromise.

If you’re like me, you’ve been enthralled with the recent story, expertly written by Sean Hollister at The Verge, about how Sammy Azdoufal built a remote control for his DJI Romo vacuum with a PlayStation controller, and ended up in control of 7,000+ robovacs all over the world. On the surface, it sounds like vibe coding gone slightly sideways. I mean, really, what could a vacuum possibly do? Turns out… a lot.

In my previous piece, "The Agentic AI Governance Blind Spot," I laid out what I believe is one of the most critical gaps in the AI governance landscape today: the three most cited frameworks in AI governance, NIST AI RMF, ISO 42001, and the EU AI Act, don’t contain a single mention of agentic AI. Not one reference to autonomous agents, multi-agent systems, or AI that takes actions with real-world consequences. The response to that piece confirmed what I suspected.

OpenClaw exposes a fundamental misalignment between how traditional enterprise security is designed and how AI agents actually operate. As an AI agent assistant, OpenClaw operates with human permissions, executes actions autonomously, and processes untrusted content as input, all while sitting outside the visibility of conventional security tools.

Approach any security, technology and business leader and they will stress the importance of governance to you. It’s a concept echoed across board conversations, among business and technology executives and of course within our own echo chamber of cybersecurity as well. For example, the U.S. Cybersecurity Information Security Agency (CISA) has a page dedicated to Cybersecurity Governance, which they define as.