Large language models are fundamentally dangerous to psychological and physical safety. The belief that ML companies will ensure AI alignment with human interests is naive. For every "friendly" model created, an equivalently powerful "evil" version becomes possible. LLMs represent a security nightmare, enabling sophisticated attacks at unprecedented scales.

Current alignment efforts fail because LLMs are chaotic systems we don't fully understand. The industry's four supposed protective moats—hardware access, secrecy, training data, and human oversight—are crumbling. Microsoft, Oracle, and Amazon are making ML training widely accessible, while Meta trained models using pirated books and internet scraping.

The lethal trifecta warns against giving LLMs untrusted content, private data access, and external communication capabilities. Yet products like OpenClaw connect LLMs to inboxes and browsers, while Moltbook creates a social network for AI agents to exchange untrusted content automatically—exactly the dangerous behavior security experts have warned against.