simonwillison.netSimon Willison's essay naming the three capabilities that make an AI agent attackable: untrusted input, sensitive access, and state change or external communication. The game's first five missions are built on this framework.The Lethal Trifecta
ai.meta.comMeta's engineering guidance introducing the Agents Rule of Two: an agent can hold at most two Trifecta capabilities on its own, and reaching for the third requires a control in between. Missions six to ten teach that separation hands-on.Practical AI Agent Security
genai.owasp.orgThe OWASP GenAI Top 10 entry on prompt injection, the technique attackers use to smuggle instructions into agent inputs. The game's forged tickets and poisoned reports are simplified versions of these cases.LLM01: Prompt Injection
nist.govNIST's framework for measuring and managing AI risk across its lifecycle. Useful for turning what you practiced here into an organizational program.AI Risk Management Framework
Accidental CISO is a narrative-driven, choice-based cybersecurity survival simulation game, by the same author.View Accidental CISO