Claude Mythos and misguided open-weight fearmongering

Anthropic's new Claude Mythos model landed this week with cybersecurity abilities strong enough that a familiar chorus started up almost immediately: open-weight versions of this thing will be catastrophic, our digital infrastructure can't handle it, ban the open stuff before it's too late. Interconnects pushes back hard on that framing, and not because the underlying capability is fake. It's because the argument collapses a pile of unknowns into a tidy policy conclusion that might actually leave us less prepared, not more.

This isn't the first rodeo. OpenAI held back GPT-2 weights in 2019 over doom predictions that never fully materialized, and GPT-4's 2023 release triggered similar bio-risk hysteria that mostly fizzled. The mistake, as the piece lays out, is treating the open-closed capability gap as frozen in place and then linking general open-weight viability to one narrow, scary use case. The writer has argued elsewhere that top open models will keep trailing frontier closed ones — and that's fine, even healthy. A 6-to-18-month lag between a closed lab shipping a capability and an open reproduction showing up gives the world time to monitor and adapt without killing the open ecosystem outright.

Where this case differs from GPT-4-era panic is that cybersecurity risk feels less hypothetical than bio-risk ever did. Fair enough. But the actual threat depends on things nobody outside Anthropic currently knows: how big Mythos really is, what tools and harness surround it, and how much compute it takes to run. Interconnects ballparks it at maybe 2x the parameters of Opus with a much less efficient serving setup — something in GPT-4.5 territory, but actually trained well this time. Serving an 8-trillion-parameter MoE model plausibly needs on the order of 100 H100s, roughly $10,000 a day. That is not laptop-and-a-dream territory. Very few actors can afford to run something like that at scale, even if weights eventually leaked or got open-sourced.

The coding angle matters too. Cybersecurity skill likely rides heavily on coding ability, and coding is one domain where open models can train on mountains of public GitHub data and stay closer to frontier performance than in messier fields like law or medicine. That's actually a reason for cautious optimism about narrow-domain competitiveness, not blind panic. The piece also flips the usual argument: if defenders get access to open models near Mythos's level, they could fine-tune them to harden their own systems, rather than only attackers benefiting.

The bigger point is about who gets to decide. Relying entirely on one private company to gatekeep the security assessment of critical global infrastructure isn't a stable arrangement either. The author isn't saying cybersecurity concerns are fake — deepfakes already crossed a similar red line in image models, so it's not like these worries are baseless. But blanket bans on open weights don't erase the technology, they just hand influence over it to whichever country keeps building it anyway.