Slide 18 of 28
Part 4 — PreventionSlide 18
PART 4
Prevention
Slides 18–28 · Seven mitigations, a matrix, a quiz, and what to do next
Slide 18 · Prevention Overview
Why defense is hard — and why it's still very possible
The bad news first. Then the strategy.
Why traditional defenses don't work here

Firewalls check source and destination. IAM policies check who is calling. EDR looks for known-bad binaries and behaviors. None of these check whether the agent's intent behind a tool call matches what it was supposed to be doing. You can have perfect network security and still have a complete tool misuse attack succeed.

The seven-layer defense strategy

MIT 01 — Least Privilege for Tools: Every tool gets only the access it needs for its specific purpose. Not the full API — just the scope this agent, for this task, legitimately requires.

MIT 02 — Action-Level Authentication: High-impact tool calls require explicit confirmation — either from a human or from a policy engine that verifies intent at invocation time.

MIT 03 — Execution Sandboxes: Tool calls that generate code or make external requests run in isolated environments with outbound allowlists.

MIT 04 — Intent Gate (Policy Enforcement): A middleware layer that validates the intent behind a tool call before it executes. Does this call make sense in context?

MIT 05 — Tool Budgeting: Usage ceilings — rate limits, call counts, cost caps — that automatically halt the agent when something looks like a loop or runaway invocation.

MIT 06 — Just-in-Time Access: Credentials and API tokens issued per-task and revoked immediately after use, so there's no standing access for attackers to abuse.

MIT 07 — Semantic Validation + Monitoring: Enforcing fully qualified tool names, detecting anomalous sequences, and logging everything with enough context to reconstruct intent post-incident.

If you can only implement two

Start with MIT 01 (Least Privilege) and MIT 02 (Action-Level Authentication). Together they address the most common root cause (over-privileged tools) and the highest-impact scenarios (destructive or financial actions). The others add depth.

← Back Let's go → Mitigation 1: Least Privilege for Tools