Anthropic says their latest model Claude Mythos is too dangerous to release publicly. It reminds me of 2019 when GPT-2 was “too dangerous” to release.

OpenAI said GPT-2 could create fake content at scale and support similar campaigns to the Cambridge Analytica scandal that influenced the election.

Today we have 1 Trillion parameter LLMs and Agents that can do so much more than GPT-2 ever could. People are running them on their machines with full write permissions and internet access.

Anthropic’s decision felt like another hype train but looking deeper into it, I see this to be a credible concern if the claims are true.

Claude Mythos has already discovered thousands of high-severity zero-day vulnerabilities in every major operating system and web browser.

To quote a video from Project Glasswing:

“On Linux…as a user with no permissions I could elevate myself to be administrator by just running some binary on my machine”

They plan to work with software maintainers to discover and patch exploits before public release. After release, it would be a tool developers can use to discover exploits before releasing new software.