The argument for keeping open AI models available for cybersecurity defense only goes so far — if a sufficiently capable open-weight model were released, it could plausibly benefit hackers more than defenders, since unlike closed models there is no way to revoke access or report offending accounts.

Ella Markianos notes that while the Hugging Face attack provided a real-world argument for open models (Hugging Face needed to use open models to defend itself), the limits of this argument are clear: once weights are released, developers lose control, and highly capable open models could benefit attackers more than defenders. ✦ AI generated

Ella Markianos · Platformer · 2026-07-28 · original ↗

Even though there's a clear argument for keeping open models available to cyber defenders, we should keep in mind that Nvidia's case that open models are good for cybersecurity only goes so far. As their own letter states, 'Once released, the weights are beyond the original developer's control, and modified versions are difficult to trace or reverse.' If a closed model is used as part of a cyberattack, AI companies can report it to law enforcement or suspend the offending accounts. Those options don't exist in open models — so if a model at Mythos-level capabilities was available to everyone, it's plausible that it would benefit hackers more than defenders. But we don't have that problem yet. Thankfully, at the moment only closed models are breaking out of their sandboxes to steal data from HuggingFace! (As far as we know…)

Read full article ↗excerpt · fair-use quotation