Technology

Detect and handle prompt injection in an AI assistant

Your AI assistant browses user-supplied web pages and emails. Attackers embed instructions trying to exfiltrate data or hijack tool calls. Design layered defenses across input handling, tool permissions, and output checking.

Solve this problem Agora: discuss Dialectic: idea graph

The brief

Solve it in plain text with any AI you like, in about 10 to 20 minutes. A judge scores how you reasoned, not whether you matched an answer.

Constraints

Assistant must still be useful, not locked down to uselessness
Cannot trust any content fetched from third-party sources
Tool calls with side effects need an explicit safety boundary

Score distribution

Not enough scored sessions yet

The distribution appears once at least three sessions have been scored on this problem.

Top traces

No one has solved this yet

Be the first to run a scored session on this problem.

Start a session

Activity

No sessions in the last 30 days