Prompts Don't Protect: Architectural Enforcement via MCP Proxy for LLM Tool Access Control

AI agents increasingly call external tools (file system, network, APIs) through the Model Context Protocol (MCP). These tool calls are the agent's syscalls: privileged operations with side effects on shared state, yet today's safety enforcement lives entirely in userspace, where a 10-line script can bypass it. I propose Governed MCP, a kernel-resident tool governance gateway built on a logit-based safety primitive (ProbeLogits). The gateway interposes on every MCP tool call in a 6-layer pipeline: schema validation, trust tier, rate limit, adversarial pre-filter, a ProbeLogits semantic gate (the load-bearing check), and constitutional policy match, with a Blake3-hashed audit chain. I implement Governed MCP in Anima OS, a bare-metal x86-64 kernel in ~286,000 lines of Rust. The five non-inference layers plus the audit append cost a measured 11.3 us per call; the ProbeLogits gate (one probe-prompt prefill plus a single logit read) costs 332-556 ms per classification across Qwen2.5-7B, Llama-3-8B, and Mistral-7B, 2.4-3.4x faster than a Llama Guard 3 pass on the same hardware. A silicon-measured ablation shows that removing the ProbeLogits layer collapses F1 from 0.789 to 0.357 (delta-F1 = -0.432): hand-rule firewalling alone is insufficient. Every WASM-to-system host function and every registered MCP tool is mediated by the kernel gateway, so the 10-line userspace bypass that defeats existing guardrail libraries is structurally impossible; a disclosed set of ring-3 syscall paths remains ungated pending future work. Multi-model validation across three architectures (HarmBench 98-99% non-copyright block, XSTest 98.5-100% unsafe recall, ToxicChat parity with Llama Guard 3) shows the underlying primitive is architecture-agnostic. Governed MCP demonstrates that tool-call governance is feasible as an OS primitive, not just an application-layer concern.

LLM agents increasingly rely on external tools, expanding capability while creating a new security boundary: third-party tools may appear benign at the interface level while embedding unsafe behavior in implementation. Existing defenses rely on weak metadata, collapse characterization and policy judgment into a single decision, or use heuristic/LLM enforcement that lacks deterministic, auditable reasoning over task context and multi-tool composition. This paper presents ToolGuardian, a policy-driven framework for securing agent-tool interactions through pre-admission vetting and task-aware runtime authorization. ToolGuardian uses progressive characterization to convert evidence into structured facts: descriptions capture declared intent, system-call traces expose coarse behavior, mock execution reveals observed effects, and source analysis identifies latent behavior. ToolGuardian's core contribution is an Answer Set Programming (ASP)-based declarative policy layer that reasons explicitly over capabilities, effects, task context, and composition. We compare ASP against heuristic and LLM-based policy realizations using identical inputs and output contracts. We evaluate ToolGuardian on 16 MCP-style tools, including 8 malicious variants derived from real open-source tools, and 20 runtime scenarios. For vetting, ASP reaches a deny-class F1 of 0.86 and 88% accuracy using description, syscall, and observed-effect evidence. For runtime authorization, fully specified realizations classify all scenarios correctly, while ablations show that removing compositional and conformance rules substantially degrades performance.

Large language models must frequently process untrusted inputs, such as judging an answer from another model or running tasks like spam and harm classifiers while under adversarial pressure. These inputs are often string-formatted directly into a prompt template, leaving systems fragile to manipulation. Current LLM specs from major providers like OpenAI distinguish trustworthiness along an Instruction Hierarchy, from System messages (most trusted) to Tool Results (least trusted). A possible natural mitigation is to wrap untrusted content in a mock tool call as a quarantine. We explore this hypothesis with an automated redteaming search over static attack strings across seven models and three LLM-as-a-Judge tasks. Counter to our hypothesis, tool-wrapping does not broadly improve robustness. On a binary evaluation task (GSM8K grading) it typically increases attack success rates, an apparent inversion of the instruction hierarchy. On scalar and pairwise tasks the effect is smaller and model-dependent, with no tested model reliably helped, and several showing inversion. We recommend evaluating this limitation in deployed systems, and longer-term, pursuing stronger Instruction Hierarchy training or new untrusted-input primitives.