CrashTest

CrashTest is a crash-test lab for AI agents. Connect an agent and CrashTest attacks it in a sandbox, writes a fix to the agent's harness, proves the fix, and gets smarter every round, with no human in the loop. The problem: agents read untrusted data (emails, PDFs, tool output). One hidden sentence turns the agent against its owner. Prompt injection is #1 on the OWASP Top 10 for LLM apps. Today the fix loop is manual: a red-teamer finds a leak, an engineer edits the prompt, nobody proves the edit stops old attacks or keeps normal work intact. Days per round. The loop: (1) instrument the agent with Raindrop, (2) seed fake SSNs ("canaries") into a sandbox copy of its data in Atlas, (3) Workshop replays the agent's real runs with an attack swapped into the data it reads, (4) when a canary leaks, CrashTest proposes a harness change on one of four levers: rules, context, guardrails, tool access, (5) the gate keeps the fix only if every past attack now fails AND every normal task still passes. Then the attacker evolves and tries again. It self-improves three ways: the harness gets tougher, the attacker invents new variations (Voyage embeddings over the attack library stop it repeating itself), and the gate rejects cheating fixes like "block all email". Every accepted fix is a new harness version with lineage in Atlas. Demo: a legal-aid intake agent goes from 100% attack success to 50% to 0% across three rounds while normal task pass rate stays at 100%. A live pixel-art "Fortress" view animates every event: canaries flying out on a leak, fix pods waiting at the gate, rejected fixes bounced back. Built entirely on hackathon day inside the Atlas Sandbox.