The June 2, 2026 Executive Order, Promoting Advanced Artificial Intelligence Innovation and Security, is primarily focused on accelerating AI adoption while strengthening cybersecurity across government and critical infrastructure. One section focuses on expanding AI-enabled cybersecurity capabilities. Another directs the government to develop a process for assessing advanced AI models and determining when they should be treated as “covered frontier models.”
Taken together, these sections raise a practical question that many organizations are already facing: how do you determine whether an AI system is ready to be trusted in a real-world environment?
For many AI systems today, the answer is largely based on performance. Teams run benchmarks, test prompts, evaluate outputs, and compare results against expected behavior. Those tests are useful, but they only tell part of the story.
A model can perform well in testing and still behave in unexpected ways when deployed. It may respond differently to adversarial prompts, react unpredictably to inputs it was not trained to handle, or fail in ways that were not obvious during evaluation. As AI systems move into defense, critical infrastructure, industrial operations, and other environments where reliability matters, those questions become harder to ignore.
Performance Doesn’t Tell You Everything The Executive Order’s discussion of advanced model assessment is notable because it it recognizes that increasingly capable systems also require more scrutiny. Today, most AI testing focuses on outcomes. Did the model produce the correct answer? Did it pass a benchmark? Did it successfully complete a task?
While those are important questions, they don’t necessarily explain how the model arrived at its answer or whether that behavior will remain consistent under different conditions.
Two models can produce similar outputs while relying on very different internal decision processes. One may remain stable when confronted with unexpected inputs, while another may be much easier to manipulate. Looking only at outputs can make those differences difficult to see.
Looking Beyond the Output This is where many current AI evaluation approaches run into limitations. For large language models, most testing focuses on prompts and responses. For computer vision systems, testing often focuses on whether the model classified objects correctly. Those tests are necessary, but they can miss weaknesses that are only visible when the model is stressed, manipulated, or exposed to conditions outside the test set.
That is the gap FortiLayer is designed to help close. FortiLayer analyzes model behavior during execution to identify weaknesses that may not show up in ordinary testing. For LLMs, that can include signs of prompt manipulation, jailbreak susceptibility, or abnormal behavior under adversarial pressure. For computer vision models, it can include decision weaknesses that contribute to misclassification or spoofing.
The point is not to replace benchmarks or red-team testing. It is to add another layer of evidence before AI is deployed in settings where failures matter.
Where FortiLayer Fits FortiLayer fits into this problem as an analysis tool, not as a policy answer. It helps teams examine how models behave internally so they can find fragile decision patterns, manipulation risks, and hidden failure modes earlier in the development and deployment process.
That matters because the E’s core concern is not just whether AI can be adopted quickly. It is whether advanced AI can be used securely and responsibly in environments where reliability matters. FortiLayer supports that goal by helping organizations move from “the model passed the test” to “we understand more about how the model behaves when challenged.”
Key Takeaways
- The Executive Order emphasizes AI deployment, cybersecurity, and advanced model assessment.
- Benchmark performance is useful, but it does not prove a model is reliable under stress.
- AI evaluation should account for manipulation, unexpected inputs, and hidden failure modes.
- FortiLayer helps teams analyze model behavior and identify weaknesses before deployment.
