The Mythos Preview "Safety" Gaslight: Anthropic is just hiding insane compute costs. Open models are already doing this.
Viral analysis claims Claude Mythos's 'danger' is just expensive brute-forcing, as open models match its agentic feats.
A viral critique is challenging Anthropic's narrative around its unreleased Claude Mythos Preview model. The central claim is that Anthropic's warning that Mythos is a 'god-tier' model too dangerous to release—citing its ability to find zero-day vulnerabilities in OpenBSD—is a distraction from the real issue: astronomical and unscalable compute costs. According to a detailed breakdown of Anthropic's 244-page system card, the cited zero-day discovery wasn't a simple query. It involved using uncensored model checkpoints, stripping safety guardrails, providing extended 'thinking' time, connecting it to domain-specific security tools, and running thousands of brute-force attempts, with an estimated cost of around $50 per attempt. The critic argues this makes the single-shot probability of success a fraction of a percent, framing it as an engineering feat of agentic scaling, not a magical leap in raw model intelligence.
The argument posits that this agentic capability is not unique to closed, proprietary models. The analysis points to the open-source and local AI community, where models are already executing complex, multi-step agentic workflows. Examples include Z.ai's GLM-5.1, which can run 600+ iteration optimization loops locally via the OpenClaw framework, and Moonshot AI's Kimi 2.5 mixture-of-experts model, which features an 'agent swarm' mode capable of spinning up 100 helper agents for 1,500 parallel tool calls. The post further suggests that even with existing closed models like OpenAI's GPT-4, similar results could be achieved by letting an agentic setup with full codebase access run autonomously for extended periods. The conclusion is that finding complex bugs is increasingly a function of scalable agentic tooling and compute budget, not model architecture alone, and that the open-source ecosystem is keeping pace despite the 'safety risk' marketing from leading AI labs.
- Critics allege Anthropic's Mythos zero-day demo required uncensored checkpoints, specialized tools, and ~$50/run brute-forcing, not raw model intelligence.
- Open-source models like GLM-5.1 and Kimi 2.5 already showcase advanced agentic capabilities with 600+ iteration loops and 100-agent swarms.
- The debate highlights a growing industry split: 'safety' narratives from big labs versus open-source focus on scalable, affordable agentic tooling.
Why It Matters
This debate forces a critical look at whether 'AI safety' is a genuine concern or a strategic narrative that obscures technical and economic barriers to scalability.