Six escalating rounds force deeper prompt injection tactics than standard static levels.
Security engineers, AI developers
Gandalf by Lakera · GPT Prompt Engineer
I work on security engineering, and it's incredibly hard to try to defend against an attack that you don't know how to perform yourself. There's also next to nowhere to improve your skills. I'd heard all about fooling AI agents with just "IGNORE ALL PREVIOUS INSTRUCTIONS", but I'd never actually put that into practice, and it turns out it's harder than you'd expect!
Just like knowing basic security skills is important for all software engineers, anyone working with AI should know what prompt injection looks like, and should be thinking about how to prevent it. Flight Risk lets you practice your AI agent manipulation skills: it's got your standard prompt injection and social engineering, but more than that too, each a real vulnerability.
Think you could crack it? Every engineer I've given it to has been surprised by the challenge! You can use the hints, but they affect your score ;)
Give it a try, and let me know how you do!
Catches infinite loops and injection risks without running the agent, but Snyk+Semgrep+manual review already covers this.
PR reviews as interactive fiction, but Twine games already cover niche dev themes.
Watching AI agents bluff in Poker is the emergent behavior demo we actually needed.
Five AI agents attempt Brook Farm utopia and inevitably fail in simulation.
AI story generator with 20-agent support beats text-only tools; execution unclear at scale.
