CVE-2026-19045

CVE-2026-19045 is a command injection in the SecretDialog component of LudisMCP, targeting the Description field that accompanies credential collection. The sanitize() method exists but fails to properly escape or neutralize shell metacharacters in this field, allowing an attacker to inject commands that execute when the credential is processed downstream. If you're running this code, the immediate priority is audit any deployment where external data can pre-populate the Description field before it reaches the user. This isn't theoretical — MCP agents and automation frameworks routinely populate metadata fields programmatically, meaning the Description field may contain context extracted from untrusted sources rather than direct user input. Treat every string flowing through SecretDialog as potentially hostile, regardless of whether it originated from a human typing or an agent's reasoning chain. The project appears unresponsive to the vulnerability report, which changes the remediation calculus. Unlike a CVE with an available patch, you're looking at either isolation (run the component in a sandboxed environment with no downstream command execution), removal (eliminate the dependency if possible), or fork-and-fix (the sanitize() method needs explicit shell-escaping logic for the Description field before it passes to any exec/popen call). The CVSS 5.3 reflects the vulnerability in isolation, but the unmaintained status means the exposure window is effectively unbounded — there's no upstream fix coming. Monitor your artifact inventories for any forks or derivative works of this library, which is a common vector when maintainers go silent. The pattern here — credential collection UI treating metadata fields as trusted — has appeared in password managers and SSH agents before, so treat this as a canary for similar design gaps in other credential-handling codebases.

Reviewed through automated stages and approved by a human before publication.