“Too dangerous to release” or just too expensive?
Claims that Anthropic’s new “Mythos” AI model is “too dangerous to release” are being weighed against more mundane explanations such as high inference costs, limited compute capacity, and IPO-era marketing incentives. Commenters point to mixed real‑world results—like modest vulnerability findings in well‑audited projects such as cURL—arguing Mythos may be only incrementally better than existing frontier models, even as some large firms report dramatic gains in bug discovery. The thread raises broader concerns about corporate control over powerful cybersecurity tools, the potential for safety rhetoric to entrench commercial moats or restrict open models, and the lack of transparent, verifiable evidence about Mythos’s true capabilities.
“Too dangerous to release” vs marketing narrative
- Many commenters see the “too dangerous to release” framing as a recycled marketing move (compared to GPT‑2/3 era messaging).
- Some argue it conveniently hides either modest capability gains or uneconomical serving costs while preserving a mystique.
- Others think Anthropic leadership likely does believe in substantial cyber risk, even if the messaging is overdramatized.
Cost, compute, and business incentives
- Strong view that Anthropic is compute‑constrained and Mythos is expensive to run; safety is seen as a useful excuse to limit access.
- Pricing details (Mythos vs Opus, preview pricing, free credits) lead some to doubt it’s vastly superior; others note larger models naturally cost more.
- Several see the timing and framing as IPO / enterprise‑sales driven hype and a way to lock in large contracts.
- There is suspicion about protecting IP and slowing competitors’ ability to distill or train on Mythos outputs.
How capable is Mythos really?
- On cyber vulns, views diverge sharply:
- One camp says Mythos is only incrementally better than other frontier models; evidence cited includes limited new findings on well‑audited projects (e.g., curl) and similar performance from other LLMs given enough compute.
- Another camp reports “revolutionary” results on very large proprietary codebases, with thousands of real bugs and design flaws uncovered, far beyond prior tools.
- Some emphasize that even small gains in security‑bug discovery could materially change offensive capabilities.
- There’s debate whether Mythos is just “a bigger model on the scaling curve” versus a meaningful qualitative shift; unclear from public data.
Safety, risk, and governance
- Concerns span offensive cyber use, possible bio‑risk, and the long‑tail of increasingly capable systems.
- Others counter that bioweapon barriers are dominated by regulation, logistics, and deterrence, not LLM access.
- Worries are raised that “safety” rhetoric could be weaponized against open‑weight models and used to entrench proprietary moats.
- A commenter from Anthropic states the bottleneck is safeguards for offensive cyber risks, not compute, and that Mythos‑class models are intended for broader deployment once controls exist; skeptics question what concrete safeguards mean.
Meta: quality of the article and ecosystem
- Many criticize the linked article as verbose, derivative, and likely LLM‑generated “slop,” though some found its cost‑focus illuminating.
- The thread also touches on HN “hug of death,” fragile WordPress hosting, and general fatigue with AI‑driven content and marketing.