Practical boundaries for models, retrieved data, agents and AI-assisted development.
A practical 2026 lab design for defending AI assistants: isolate untrusted content, constrain tools, bind approvals to actions and measure policy failures.
Create a versioned AI evaluation set with task outcomes, refusal cases, retrieval permissions, human review and explicit release gates beyond leaderboard scores.
Review AI-assisted changes through authorization tests, dependency checks, secret handling and deployment evidence rather than accepting plausible-looking code.
Separate model identity, serialization safety and execution permissions. Build a controlled acceptance process for weights, tokenizers and model repositories.
Evaluate browser and on-device AI by mapping model downloads, inference, telemetry, synchronization and cloud fallback instead of trusting a local-AI label.
Protect payment changes from voice, video and email impersonation with known-channel verification, separation of duties and a practical tabletop exercise.
Plan for changing threats without betting on a prediction. Map long-lived dependencies, choose actions that help today and record the evidence that would change your roadmap.
