Defence, red teaming or specialised access: verified security teams are granted less restricted access. Glasswing claims to have identified over 129,000 confirmed vulnerabilities.
Anthropic is expanding its Cyber Verification Program (CVP), which gives verified security professionals access to its models with lightweight cyber filters, and integrating Project Glasswing, which until now has been reserved for a few organizations protecting critical software with Claude Mythos. The three tiers provide access to the company's most capable models, including Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. Consumer models, on the other hand, retain cautious guardrails that block most offensive uses.
Three levels, three degrees of control
Defense access covers incident response, malware analysis, and vulnerability validation; it is aimed broadly, from regional hospitals and municipal utilities to open source maintainers, with a response promised in a matter of days. Red Team access adds authorized penetration testing, on a multi-week file and for organizations only; actions that could cause physical damage or massive disruption remain blocked in real time. Specialized access, the least filtered, is reserved for a small number of actors authorized to test sensitive systems: avionics, power grids, telecoms, interbank transfers, administrative networks. Data must be kept to monitor abuse, pending a retention-free option planned for this fall.
Anthropic also quantifies the effect of Glasswing: between April and July 2026, its partners confirmed at least 129,000 vulnerabilities with Mythos, in addition to 5,500 flaws found by the company in open source projects from April to October. More than 33,000 were classified as critical or of high severity. The company estimates the real impact to be at least five times higher, with the count covering only some of the partners.
Read the full article on the original source
Go to the source →