Matryoshka Rlm MCP Server

Tool "nucleus_execute" appears to run shell commands or evaluate code (parameter "command"). Arbitrary execution driven by model input is one of the most dangerous MCP capabilities; combined with any untrusted input it becomes RCE.

Fix: Sandbox execution, allowlist commands/arguments, and never pass model output to a shell unescaped.

Location: tool nucleus_execute

In the server's implementation (`dist/constraints/verifier.js:263`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: tion const fn = new Function("result", `"use strict"; return (${invariant});`); return Boolean(fn(result

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/constraints/verifier.js

In the server's implementation (`dist/logic/synthesis-integrator.js:696`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: const fn = new Function("input", `return ${generatedCode}`); return {

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/logic/synthesis-integrator.js

In the server's implementation (`dist/persistence/predicate-compiler.js:170`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: eval return new Function("item", `"use strict"; return (${code});`); } catch (e) {

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/persistence/predicate-compiler.js

In the server's implementation (`dist/sandbox.js:347`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ('\\n'); } `; vm.runInContext(initCode, vmContext); return { async execute(code, timeoutMs = 30000

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/sandbox.js

In the server's implementation (`dist/synthesis/coordinator.js:64`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ); } const fn = new Function("return " + code)(); if (typeof fn !== "function") { throw new Error("S

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/coordinator.js

In the server's implementation (`dist/synthesis/evalo/compile.js:107`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: neration return new Function("input", `return ${code}`); } catch (err) { throw new Error(`Failed

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/evalo/compile.js

In the server's implementation (`dist/synthesis/evolutionary.js:219`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: const fn = new Function("return " + code)(); return examples.every((e) => { con

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/evolutionary.js

In the server's implementation (`dist/synthesis/relational/engine.js:241`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: const fn = new Function("input", `return ${code}`); const result = fn(input);

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/relational/engine.js

In the server's implementation (`dist/synthesis/relational/interpreter.js:148`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ry { const fn = new Function("input", `return ${code}`); return fn(input); } catch { ret

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/relational/interpreter.js

In the server's implementation (`dist/logic/synthesis-integrator.js:687`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ion\b/, /\bfetch\b/, /\bchild_process\b/, /\bReflect\b/, /\bProxy\b/, /\bwith\b/, /\bdelete\

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/logic/synthesis-integrator.js

In the server's implementation (`dist/persistence/session-db.js:74`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: y returned rows. */ exec(sql) { this.db.exec(sql); } /** Execute a single SQL statement, ignoring an

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/persistence/session-db.js

In the server's implementation (`dist/synthesis/coordinator.js:24`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ion\b/, /\bfetch\b/, /\bchild_process\b/, /\bReflect\b/, /\bProxy\b/, /\barguments\b/, /\bwindow\b/, /\b

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/coordinator.js

In the server's implementation (`dist/synthesis/evolutionary.js:198`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ion\b/, /\bfetch\b/, /\bchild_process\b/, /\bReflect\b/, /\bProxy\b/, /\barguments\b/, /\bwindow

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/evolutionary.js

In the server's implementation (`dist/synthesis/relational/interpreter.js:127`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ion\b/, /\bfetch\b/, /\bchild_process\b/, /\bReflect\b/, /\bProxy\b/, /\barguments\b/, /\bwith\b/, /\bdelete\b/,

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/synthesis/relational/interpreter.js

Tool "nucleus_execute" takes a command-shaped parameter "command" with no enum/pattern constraint. Free-form, model- or attacker-controlled arguments reaching a shell is the command-injection precondition.

Fix: Constrain the parameter (enum/pattern), or build the command from a fixed template with escaped args.

Location: tool nucleus_execute · inputSchema.properties.command

In the server's implementation (`dist/treesitter/parser-registry.js:109`): Loading a module chosen at runtime (from a variable) can pull in and run attacker-influenced code paths. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: grammarModule = require(config.package); } catch (err) { throw new Error(`Grammar pa

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/treesitter/parser-registry.js

Tool "nucleus_execute" can mutate/egress but declares no destructiveHint. Clients that don't default to spec-safe behavior may not prompt before running it.

Fix: Declare accurate annotations, and gate destructive tools on user confirmation regardless.

Location: tool nucleus_execute