agent-goal-hijack-defender

What problem does it solve?

This Skill protects a multi-step AI agent from having its objective quietly changed by untrusted content during execution, preventing scope creep, task substitution, and plan drift.

Core Features & Use Cases

• Pinned Goal Record: Defines an explicit objective, scope, constraints, and authorizing principal outside the model context so the run can be checked against a stable source of truth.
• Step Tracing and Deviation Checks: Validates each plan or replan step against the pinned goal using deterministic checks where possible and weaker model-mediated checks when necessary.
• Mutation Control and Drift Response: Restricts goal changes to an authenticated principal, detects hijack signals such as scope expansion or replan storms, and routes responses to halt, re-ground, degrade, or contain.
• Use Case: An agent browsing documents starts drifting toward an attacker’s requested target; this Skill keeps the original goal intact and flags the deviation before side effects accumulate.

Quick Start

Use the agent-goal-hijack-defender skill to design a pinned goal record, step-tracing checks, and drift responses for a multi-step agent that must ignore untrusted content trying to redirect its objective.