CVE-2026-54561: mcp-memory-keeper: Arbitrary local file read in context_import via unvalidated filePath

context_import passed the caller-supplied filePath directly to fs.readFileSync with no path confinement. A malicious MCP client — or an LLM agent that is prompt-injected into calling the tool — could point filePath at any file readable by the server process, outside any session or export directory:

• Full disclosure (JSON files): a valid-JSON target (e.g. another user’s exported session, or a *.json credential / service-account file) is parsed and imported into the caller’s session, then retrievable verbatim via context_get / context_export.
• Partial disclosure (any file): for a non-JSON target (e.g. /etc/passwd, an SSH key, a .env), JSON.parse throws and V8 includes a snippet of the file’s leading bytes in the SyntaxError message, which was returned verbatim to the caller.

Both ../ traversal and absolute paths worked — there was no path confinement of any kind.

In a typical MCP deployment the server runs on the developer’s machine, so the reachable set includes other users’ exported memory sessions, JSON credential/config files, and (in leading-bytes form) .env files, SSH keys, and /etc/passwd. The trigger is a tool argument, so the realistic threat model is an LLM agent prompt-injected into calling context_import, or any MCP client connected to the server.