I broke my own governed MCP server by hand, then built the scanner that catches the class

Short summary

A support role in Warden couldn't see billing tiers but could filter on them, leaking values through result counts. To catch such runtime authorization bypasses, the author built Siege: a differential security scanner that exercises MCP servers as different roles and diffs outputs to detect access-control leaks that static scanners miss.

• •Discovered a filter-input vulnerability in Warden, an MCP access-control layer
• •Built Siege to detect runtime authorization breaches via differential testing across roles
• •Identifies four classes of access-control vulnerabilities that static manifest scanners cannot catch

Generated with AI, which can make mistakes.