If you're building agents, this is worth knowing.

Simon Willison (who coined the term "prompt injection") describes three capabilities that are individually fine but devastating together, the lethal trifecta includes:

1. Access to private data (such as internal databases, emails, etc.)

2. Exposure to untrusted content (like web pages and external text)

3. The ability to exfiltrate data (a tool or a channel of sorts that reaches the outside world)

Combine all three in one agent and you've got a perfect storm: untrusted content can hijack the agent into reading private data and then persuade the agent to send it out.

To teach the lethal trifecta, I built a Gandalf-style puzzle: 10 levels, each a real injection concept, with the full trifecta as the final level. It's free, there's no sign up and initially it's beginner-friendly.

See if you can chain all three together and complete the trifecta!

Show HN: The AI Lethal Trifecta
getjailbroken.com