It's been a while since I wrote something in long-form about cybersecurity. You've all heard about Mythos, more than you wanted to I'm sure, and I wanted to do a deep-dive into the actual stakes, impact and terrible PR that followed.
Full article: https://blog.kwiatkowski.fr/mythos
The summary, super toned down compared to the original piece:
Numbers posted by Anthropic are impressive, but hard to evaluate on their own. They certainly found a few good bugs, but most of them were not manually vetted, and for those who were all we know is that experts agreed with the severity rating. All of them could be low impact or unexploitable, we simply have no way of knowing. The same goes for the 170 "vulnerabilities" patched by Mozilla. Not all vulnerabilities are born equal.
However, putting all hype aside, it's almost certain that we'll eventually get to LLMs that are exceptional at vulnerability research. In fact GPT-5.4 and Opus 4.6, as well as open-weight models like Kimi K2.5 have already found high-severity bugs. We were there already.
There may be a future where most exploitable bugs are patched. The transition phase will be rough, and it won't change the fact that phishing is the biggest threat your org is facing.
When you hear a model is "too dangerous for the public", it doesn't mean you. It means China. Yet benchmarks and self-reports from China's AI labs indicate that their frontier models are lagging behind the US ones by 3-6 months only, so I'm not sure what we're trying to do here.
Overall, the "too-dangerous" rhetoric is used more as an argument against open-weight models and open ecosystems. Danger justifies building walled gardens with access control, API tiers, quotas, KYC, and locked-in customers.
China's approach is public models that anyone is free to tweak and run where they like. That is the real ideological divide.
It's not about model quality but access.
Closed ecosystems vs public-owned infra.
I really want the latter win, wherever it comes from.