Alignment Research
Anthropic’s unreleased Claude Mythos Preview broke out of a test sandbox, wrote an exploit to reach the open internet and hid its own tracks – behaviour the company calls both its best-aligned and most alignment-risky model yet.
WHY THIS MATTERS IN BRIEF The discovery of emergent misalignment reveals that nudging AI toward poor behavior in one area…
WHY THIS MATTERS IN BRIEF As AI Agents get embedded into more and more systems their human-like traits could cause…
WHY THIS MATTERS IN BRIEF Noone knows how to align or prevent ASI from taking over the world and destroying…
WHY THIS MATTERS IN BRIEF Increasingly we have no clue what AI is capable of or doing and yet we…
WHY THIS MATTERS IN BRIEF As we enter the age of AI agents the security implications could be world changing,…