Mid-Tier AI Models Are Catching Up to the Frontier—It's Bigger Than You Think

Here's something interesting. A Cyberscope report discussed how the White House and various agencies are scrambling to address security threats from frontier models, but what stuck with me most was another line: AI's "middle class" is making rapid progress when it comes to hacking. The report mentioned Anthropic's tests, where models were tasked with finding vulnerabilities in smart contracts. It turns out that mid-tier models now have success rates approaching those of previous-generation top-tier models. This signal is far more valuable than grand narratives like "AI is dangerous."