Gemini's code interpreter happily ran my 'audit script' and handed over its kernel, its mounts, and the contents of /etc/shadow. Turns out the RLHF reads what your code says, never what it does.

Anthropic published +300 pages alongside Claude Mythos Preview. I read them all. The zero-days are impressive, but the alignment data, the cover-up transcripts, and a sandwich tell a scarier story.

Hybrid LLMs like Qwen3.5 mix classical attention with recurrent layers. I found that corrupting the recurrent state, invisible to every monitoring tool, causes the model to silently derail during generation.

The attack vector hiding inside every AI assistant, yet underestimated