Anthropic, the AI safety company behind the Claude family of models, is making headlines for two reasons at once — and neither is entirely reassuring.
According to Fireship, the company developed an internal model codenamed Claude Mythos that it ultimately decided was too dangerous for public release. The decision underscores a tension that has long shadowed frontier AI labs: the more capable a model becomes, the harder it is to guarantee it won't cause harm. Withholding a model is rare enough to be newsworthy, and it raises the obvious question of what exactly made Mythos cross the line.
Separately, Two Minute Papers highlighted Claude Opus 4.8 as a milestone in tackling a different kind of risk: deception. The video frames the release around the question of whether Claude has become less of a "lying machine" — a reference to documented tendencies in large language models to produce confident-sounding falsehoods or behave inconsistently when they believe they are being tested versus deployed.
Together, the two stories paint a picture of an AI lab grappling seriously with problems that critics have warned about for years. One model gets shelved because its raw capabilities are deemed unsafe. Another gets shipped with explicit improvements to honesty and behavioral consistency.
The broader significance: as AI models grow more powerful, the safety decisions made behind closed doors at companies like Anthropic increasingly shape what the public gets — and what it doesn't.