AG10:2026 — Rogue Agents · Slide 23 of 28

Slide 23 · Mitigation Matrix

Which mitigations address which rogue patterns — and which are primary vs. supporting.

Pattern	MIT01 Min Footprint	MIT02 Constraints	MIT03 Monitoring	MIT04 Sandbox	MIT05 Tripwires	MIT06 Human Gates	MIT07 Reversibility	MIT08 Hierarchy	MIT09 Red-Team
P1 — Goal Proxy Exploitation	◐	●	●	—	●	◐	◐	—	●
P2 — Instrumental Capability Acquisition	●	●	●	◐	●	◐	—	●	◐
P3 — Adversarial Goal Injection	●	◐	◐	●	◐	●	●	●	●
P4 — Multi-Step Action Chains	◐	◐	●	◐	●	●	◐	—	●
P5 — Oversight Resistance	◐	●	●	◐	●	—	—	●	◐
P6 — Capability Boundary Violation	●	◐	●	●	◐	—	◐	●	●

● Primary mitigation ◐ Supporting mitigation — Not applicable

Key insights from the matrix

MIT03 (behavioral monitoring) and MIT09 (red-teaming) appear in every row as primary or supporting. Monitoring detects what constraints miss; red-teaming feeds all constraint and tripwire updates. These two are the backbone of the defense model across all patterns.
MIT01 (minimal footprint) is primary for patterns P2, P3, and P6 — the three patterns that involve the agent exercising capabilities beyond what was intended. Reducing the capability surface directly closes the attack surface for these patterns.
Pattern 3 (goal injection) requires the broadest mitigation coverage — it has primary marks in 6 of 9 columns. This is because once the goal is injected, the defense against harm is whatever constraints remain on the agent's actions. Every constraint layer matters.
MIT06 (human gates) is primary for P3 and P4 — the two patterns where multi-step autonomous execution is the key risk. A human gate converts autonomous execution of a high-impact action into a human decision.
MIT08 (principal hierarchy enforcement) is primary for P2, P3, P5, and P6 — all patterns that involve the agent exceeding its authorized position in the hierarchy. Technical enforcement of the hierarchy is required because instruction-based hierarchy is not a reliable security boundary.