August 1 came and went without a word from the US government on its plan to create a secret benchmark for testing the most advanced AI models' cyberattack abilities. This benchmark was supposed to help federal agencies identify which AI systems are so powerful that they need extra oversight.
The whole effort traces back to an executive order signed in June 2026 by former President Trump. It tasked agencies like the NSA, CISA, Treasury, and NIST with building a classified framework to evaluate frontier AI models. The idea is to have a reliable way to stress-test AI systems for their potential to perform complex cyber operations. The NSA director gets the final say in labeling any AI as a "covered frontier model," triggering additional government scrutiny.
Part of the plan included a voluntary system where AI companies would give the government up to 30 days of early access before releasing their models. However, as of late July, talks with major developers OpenAI, Anthropic, Google, Microsoft, and Amazon were still unresolved. Interestingly, Meta was not part of these negotiations. Their focus on open-sourcing models like Llama doesn’t align with handing over early access to the government, which may explain their absence.
This delay raises questions about how prepared the US is to regulate AI risks just as these technologies grow more powerful and widespread. With AI's potential for cyber threats increasing, the government’s ability to enforce standards and safeguards remains unclear. This comes amid broader tensions in technology regulation, as seen in other sectors where legal battles and regulatory pauses continue to shape the space.
This article is for informational purposes and does not constitute financial advice.
