— I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model, he writes to chief compute officer Tom Brown, according to Semafor, who scooped the story.
Mythos 5 was only intended for a limited release through the Glasswing project, for use by a trusted set of government agencies and corporations to test their cyber defenses.
The note does not mention Fable 5, the first Mythos-class model slated for general release to all of Anthropic’s customers. This model had such strict safeguards it quickly became a joke in the AI community, refusing to answer anything even remotely related to security or biology.
The letter also says that «Anthropic has committed to work with the U.S. government on protocols and standards and releases,» and talks are progressing toward a release of Fable, though «the timeline is unclear,» according to Semafor.
Read more: Scoop by Semafor, comment from Anthropic, Reuters, CNBC, and Politico.