Overview
- Payward, which announced Monday, has been admitted to Anthropic’s invite-only Project Glasswing and gained restricted access to the Claude Mythos 5 cybersecurity model.
- Claude Mythos 5 can read code at machine scale, flag weaknesses, suggest fixes and help test exploit paths, but Anthropic limits access and requires 30-day safety monitoring of submitted data.
- Payward says it will run Mythos 5 across its environments in the coming weeks and route model findings into its existing security review process rather than applying automated fixes, with human validation required.
- The company will report validated vulnerabilities in third-party open-source components to their maintainers but has not disclosed first scan targets, whether production systems will be scanned, a deployment timetable, cost, or a coordinated disclosure policy.
- Project Glasswing has previously surfaced thousands of high- and critical-severity findings and the U.S. government has allowed Mythos 5 access for vetted defenders, a development that could speed patching but also raise verification burdens for open-source maintainers and security teams.