Loophole: Adversarial Agents To Stress Test Your Morality — Brendan Rappazzo, Morgan Stanley
An insurance company does not train its risk model on your DNA. It trains on artifacts derived from your DNA. Under the moral code the user wrote, that is immoral and perfectly legal, and it is the kind of gap Loophole exists to surface. Brendan Rappazzo is a machine learning researcher at Morgan Stanley, though he opens by saying the project is his own open source work, unrelated to his employer. It began with a personal question. He had sent his DNA to a consumer ancestry service, then started hearing how such samples get used in forensics and cold cases, and realized he could answer case by case whether he consented, but had no way to enumerate the cases. His reframing is that a legal system is society doing the same translation from moral belief into written rule, and that common law leans so hard on case law precisely because the nuance is hard to state up front. So Loophole generates synthetic case law against you. You describe your morals in plain language, one agent drafts them into a formal legal code, and then two adversarial agents work the code from opposite sides. One hunts for loopholes, things immoral but legal. The other hunts for overreach, things moral but illegal. A judge agent decides whether a contradiction is just a sloppy translation it can patch on its own, or a genuine hole in your morals that gets escalated to you. The branches he is exploring go further: codified system prompts for customer facing bots, contracts checked against a company's terms of service, and a Senate simulator that rewrote a Medicare bill until it cleared 52 votes. Speaker info: - https://x.com/brendanh0gan - https://www.linkedin.com/in/brendan-rappazzo-hogan-763734115/ - https://www.bhogan.net Timestamps: 0:00 - A game built on adversarial agents 1:03 - The DNA question that started it 2:44 - Morals, law, and the translation problem 4:33 - How the loop works: loophole, overreach, judge 7:15 - An overreach the judge could not patch 8:05 - Branch one: constitutions for chatbots 9:47 - Branch two: decentralized contracts 12:22 - Branch three: simulating the US Senate 14:04 - Hill climbing a bill toward passage 14:55 - Scaling to 500 personas per state
More like this

Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind

How We Solved Agent Building — Andrew Qu, Vercel

No Memory, No Harness: Why the Database Is the Last Line of Defense — Kay Malcolm, Oracle

We let an AI agent execute Bash and lived to talk about it — Sarah Sanders, PostHog
Join the discussion
Sign in to join the discussion
Sign in