Why Anthropic Believes Its Latest Model Is Too Dangerous to Release

Key Takeaways

• Mythos Preview discovered thousands of high‑severity vulnerabilities.
• Anthropic grants limited access to ~50 critical‑infrastructure companies.
• $100 million in access credits offered for security audits.
• Exploits Firefox JavaScript bugs successfully 72% of the time.
• Delay signals AI firms may keep frontier models for internal use.

Pulse Analysis

Anthropic’s Claude Mythos Preview marks a watershed moment in AI safety, illustrating how large language models can transition from passive assistants to active cyber‑offense tools. By autonomously identifying and chaining together bugs in operating systems such as OpenBSD and Linux, the model reduced the cost of sophisticated attacks to a few thousand dollars—roughly $20,000 for 1,000 runs—far below the expense of hiring elite human hackers. This capability forces enterprises to rethink traditional threat models, as AI can now surface decades‑old vulnerabilities that have evaded human discovery, prompting a surge in demand for AI‑augmented security audits.

The strategic decision to limit Mythos Preview to about 50 vetted organizations, including tech giants like Google, Microsoft, Nvidia, Amazon and Apple, reflects a defensive acceleration approach. Anthropic is allocating $100 million in access credits so these partners can harden their critical infrastructure before the model becomes widely available. By collaborating under the Project Glasswing umbrella, these firms aim to patch high‑severity bugs pre‑emptively, effectively turning a potential weapon into a catalyst for industry‑wide hardening. This partnership model may become a template for future AI rollouts, where controlled exposure balances innovation with risk mitigation.

Beyond immediate security concerns, Anthropic’s restraint signals a broader market trend: frontier AI models may increasingly stay behind corporate firewalls rather than being democratized. With revenue run‑rate now exceeding $30 billion and compute costs soaring—Mythos Preview commands $25 per million input tokens and $125 per million output tokens—the economics of mass deployment are daunting. Companies may opt to reserve their most powerful models for internal use, preserving competitive advantage while avoiding regulatory scrutiny. As AI capabilities accelerate, the industry will grapple with the tension between open innovation and the imperative to safeguard digital ecosystems.