iTnews reported that Anthropic is opening its strongest AI models to more vetted security teams. It is also merging two programs it has run for six months into one Cyber Verification Program. Glasswing gave groups protecting critical software access to Claude Mythos, its most cyber-capable model family. The older program gave vetted teams fewer safeguards on Claude Opus and Sonnet. Anthropic said Glasswing partners found at least 129,000 verified flaws from April to July. Its own open-source scanning found 5,500 more by October. More than 33,000 so far rate critical or high.
Anthropic said the totals probably undercount, since only a limited number of partners were surveyed. It expects the true impact to be at least five times higher. The new program has three tiers, each with its own checks. Defence covers incident response and malware analysis. Red Team adds authorised penetration testing, for organisations only. Specialised has the lightest limits and is for a few groups cleared to test systems such as power grids. Glasswing members move into it. iTnews says Anthropic vets "each member" with the US government. It is unclear whether that covers all tiers or only Specialised. No start date is given.
Both older programs already vetted users, so this widens access rather than adding a lock. Buyers can ask vendors which tier they hold and who vetted them. They can also weigh 33,000 serious findings, likely an undercount, against their own patch queue.