A new compliance overview from Obsidian Security marks a quiet but significant shift in how 2026's patchwork of AI rules actually bites. Earlier AI regulation focused on the model itself — training data, bias, explainability, output filtering. That's no longer where the scrutiny lands. The overview notes agentic systems add two new axes regulators weigh just as heavily: access (what an agent can reach) and action (what it does with that reach). That reframing matters because it shifts compliance from a one-time audit of model behavior to a continuous question about credentials and permissions — the kind of question that doesn't have a clean answer in most enterprise deployments right now. The number that should worry compliance teams: industry analysis cited in the piece puts roughly 90 percent of deployed agents in an over-permissioned state, meaning the access surface is almost always larger than anyone has documented. That's not a theoretical gap. It's the default condition. The instruments doing the grading are already active, not hypothetical. The EU AI Act imposes logging, human oversight, and post-market monitoring for higher-risk uses. NIST's AI Risk Management Framework applies its Govern, Map, Measure, and Manage functions directly to agentic systems, even though it's voluntary. ISO/IEC 42001 asks for a documented AI inventory and continuous monitoring. None of these were written with a non-human identity in mind — one that can inherit a builder's access and act on behalf of any invoker that calls it, as the overview puts it. That's the structural mismatch: identity governance built for people and service accounts is now being asked to cover entities that don't behave like either. What this means practically is that 'compliance' for an agentic system is no longer satisfied by a model card or a bias audit. Auditors will ask what an agent could touch, whether that access was ever reviewed, and whether there's a record of what it actually did with it. Most organizations currently can't answer the first two questions cleanly, let alone the third. The regulatory instruments exist. The documentation they require generally doesn't. This is a different failure mode than the 'half of enterprises give agents the keys' story from a few days ago — that was about provisioning. This is about the parallel fact that the rules now treat access and action as the primary compliance surface, not the model's outputs, and that most deployed agents already fail the access half of that test.