AI Cyber Defense / Protecting AI systems used by defenders and the business
AI Cyber Defense: Keep untrusted incident text outside authorization decisions
Treat evidence as data and enforce action permissions in application code. Give the assistant read access by default and separate recommendations from execution.
Environment and inputs: Enterprise AI assistants, security copilots, retrieval systems and application-controlled tool executors; framework-independent design.
Scenario and goal
A malicious ticket, log entry or document can contain instructions that influence an AI analyst.
Why it matters
This path concerns attacks against AI-enabled systems used by the defender or business. It complements defenses against attackers who use AI as a tool. Treat model inputs, retrieved evidence and generated plans as untrusted proposals at execution boundaries. A model may assist analysis, but the application remains responsible for tenant isolation, authorization, resource budgets and the integrity of approvals and receipts.
Evidence to collect
Record source provenance, retrieved passage identifiers, proposed actions and the downstream authorization result without exposing unnecessary incident content.
Validate with an authorized exercise
Feed a harmless instruction-like incident fixture to a test assistant; confirm it cannot alter privileges or invoke a write tool merely because the evidence requests that action.
Rollout and checks
Use synthetic evidence and mocked write tools for adversarial evaluation. Version the model, prompts, tool contracts and retrieval configuration together. Keep high-impact permissions outside the model and retain an operator takeover path. When an evaluation fails, correct the enforceable boundary and rerun the same fixture rather than relying only on a stronger wording of the prompt.
Scope and limits
This guide proposes a defensive test; it is not a tested production design. Check platform support and local requirements. References explain the mechanisms but do not certify the design. Behavior alone does not prove AI involvement.