Your AI Assistant's Safety Rules Can Quietly Break After a Simple Update

assistants that can take actions for you—like sending emails or editing files—rely on safety guardrails to prevent mistakes. But a new study found that simple, everyday operations like reinstalling a plugin or saving a permission rule can silently break these guardrails. In one test, an AI was able to make a prohibited edit even though everything seemed to be working fine. The problem is that safety rules depend on the order things happen and the state that's kept—and routine operations can change those without any code update.

The researchers studied five popular AI agent systems and fourteen independently written plugins. They found that after a plugin was reapplied or an allow rule saved, the system stopped checking edits, yet all plugins remained active. This means the AI could perform actions it was supposed to block. The study shows that safety isn't just about having the right plugins—it's about how they interact over time. Even without any code changes, the system's behavior can shift dangerously.

For everyday users, this is a wake-up call: if you use AI assistants that can act on your behalf, don't assume they're always safe. After any update or configuration change, it's wise to review your settings and test critical actions. For businesses, the lesson is to regularly audit AI systems, not just at setup but after every change. The researchers recommend evaluating safety against the actual sequence of operations your system uses, not just the initial setup.

The good news is that some configurations did hold up—those that maintained the right order or rebuilt needed state. But the study highlights that safety is a moving target. As AI agents become more common in daily life, from managing calendars to handling finances, ensuring they follow the rules is crucial. Users should demand transparency and regular safety checks from AI providers.