Subscribe

Stage 7 · Advanced

AI Security

Stolen tokens, poisoned tools and the other ways an AI system gets turned against its users.

You'll be able to: Spot the attack paths in an agent stack and close them.

0 of 5 done
  1. 1 Cinematic Short 47 sec Failure mode + fix Approval Fatigue When an AI coding agent asks permission for everything, approving becomes a reflex, and the one dangerous request gets the same tap as the forty safe ones.
  2. 2 Cinematic deep dive 9 min Failure mode + fix Grok Bot: Agents on One Computer Grok Bot gives every account one persistent cloud computer shared by all its bots, so one cloud computer is one trust zone: a bot that reads the open web can borrow any session another bot left behind, and a taught routine does real, irreversible work while nobody watches, unless you design for both.
  3. 3 Cinematic Short 60 sec Failure mode + fix Five Bots, One Computer Grok Bot gives your account one persistent cloud computer, and every bot you create works on it at once, sharing its files, browser sessions and logins: great for teamwork, and the reason a bot that reads a hostile page can act as you.
  4. 4 Cinematic Short 60 sec Failure mode + fix The Rule Was Only a Wish A dot is OpenAI's always-on ChatGPT agent that works on goals, not prompts. What it may do on its own is decided by layers that are either enforced (code, app permissions) or only read (Custom Rules): code and permissions hold, rules are requests.
  5. 5 Cinematic Short 60 sec Failure mode + fix Put Your Agent in a Box Approval prompts make you the security boundary, and people get tired; a sandbox lets the operating system be the boundary, so the agent can work without asking and still can't wipe your home folder or post your keys to a stranger.
  6. Coming soonAgent SandboxingAnnounced at the end of an episode in this track. Subscribe to catch it.
  7. Coming soonAgent sandboxing: how isolation really worksAnnounced at the end of an episode in this track. Subscribe to catch it.