Unauthenticated OS Command Execution in Flyto2 Core HTTP MCP
Flyto2 Core versions 2.26.2 through versions before 2.26.4 contain an authentication omission in the HTTP MCP endpoint. JSON-RPC tools/call requests can forward execute_module arguments to the module registry without the require_auth dependency enforced by REST execution routes. Missing module filtering and a default denylist that does not cover sandbox modules allow unauthenticated callers to invoke sandbox.execute_shell. Its command parameter reaches asyncio.create_subprocess_shell, enabling arbitrary OS command execution with the server process's privileges. The endpoint is locally accessible under the default loopback binding and remotely accessible when bound to a reachable non-loopback interface.
What it means and what to do now: patch path, interim mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
Successful exploitation compromises confidentiality, integrity, and availability within the server process's security context. Attackers may read or modify files, steal accessible credentials, disrupt services, and use available permissions and connectivity for lateral movement or host takeover. Root-level execution was demonstrated inside a Docker container; this does not establish escape from the container or compromise of the underlying host.
If you can’t patch tonight, do this now.
Until upgraded, disable the HTTP MCP endpoint or stop the affected API service. If continued operation is necessary, enforce authenticated access controls and network filtering. Retain loopback binding to limit remote exposure, recognizing that local processes can still exploit the endpoint. Run the service with least privilege. Defensive code changes include authenticating MCP POST and DELETE routes, passing module_filter into JSON-RPC handling, and adding sandbox.* to the module denylist.
Patch, then assume compromise.
Upgrade Flyto2 Core to version 2.26.4 or later. Verify that unauthenticated MCP requests are rejected and that module restrictions prevent unauthorized execution of dangerous modules, including sandbox.execute_shell.
No public exploits tracked yet. Mallory keeps watching.
2 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.