# Loophole: Adversarial Agents To Stress Test Your Morality — Brendan Rappazzo, Morgan Stanley

Source: [AI Engineer](https://www.youtube.com/watch?v=hOWU0KPUp1k)  
Feed7 permalink: https://feed7.dev/p/loophole-adversarial-agents-to-stress-test-your-morality-brendan-rappazz-1wjlbog  
Published: 2026-09-14T15:00:19.000Z  
Trust: Source Linked (source_linked)

## Why Included

Loophole turns a natural-language policy into rules, then uses adversarial agents to find forbidden allowances and wrongful refusals. It is a useful pattern for testing agent constitutions.

## Source Summary

Loophole converts a user's stated morals into a codified rule set. One adversary searches for immoral-but-legal cases, another finds moral-but-illegal overreach, and a judge either patches the rules or asks the user to resolve an ambiguity.

## Practical Implication

Reuse the pattern to test an agent constitution or system prompt before deployment: generate both unsafe compliance cases and excessive-refusal cases, judge them against the original intent, and preserve unresolved conflicts for human review. The project is **open source** and terminal-based.

## Agent-Ready Context

Loophole converts a user's stated morals into a codified rule set. One adversary searches for immoral-but-legal cases, another finds moral-but-illegal overreach, and a judge either patches the rules or asks the user to resolve an ambiguity.

Reuse the pattern to test an agent constitution or system prompt before deployment: generate both unsafe compliance cases and excessive-refusal cases, judge them against the original intent, and preserve unresolved conflicts for human review. The project is **open source** and terminal-based.

Its broader contract and government ideas remain exploratory. The Senate simulation and experiments using **500 synthetic personas per state** rely on model-generated values and behavior, so they should not be treated as verified representations of people or voting outcomes.

## Connected Context

Feed7 judgment across 778 accumulated Signals:

This turns constitution testing into a two-sided adversarial loop that searches for both unsafe compliance and excessive refusal, then patches rules or preserves ambiguity for human judgment. It complements trace-driven prompt evaluation with targeted counterexample generation, but model-generated adversaries and judges cannot establish that the resulting rules faithfully represent people or resolve latent multi-agent behavior.

- [How Evals and Prompts Shape Agent Behavior — Preetika Bhateja & Daniel Bump, YouTube Ads](https://feed7.dev/p/how-evals-and-prompts-shape-agent-behavior-preetika-bhateja-daniel-bump-1cmecaw) — Loophole supplies a structured way to generate the small failure sets that the candidate’s trace-review and calibrated-judge improvement loop requires.
- [The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI](https://feed7.dev/p/the-future-of-evals-from-llm-as-a-judge-to-agent-as-a-judge-aparna-dhina-1fu560o) — Its adversary-and-judge workflow is an agent-based evaluation of variable cases, but unresolved ambiguities still require the retained human-review layer.
- [The Unreasonable Effectiveness of Separating the Task from the Model — Maxime Rivest & Isaac Miller](https://feed7.dev/p/the-unreasonable-effectiveness-of-separating-the-task-from-the-model-max-0vu7lvy) — Codifying morals as a rule contract separates intended behavior from the adversaries, judge, and prompts used to test and revise its implementation.
- [What LLM Agents Say When No One Is Watching: Social Structure and Latent Objective Emergence in Multi-Agent Debates](https://feed7.dev/p/2607-02507v1-1ctgeey) — Observed divergence between agents’ public and private positions warns that Loophole’s multi-agent judgments may reflect interaction structure rather than only the encoded moral intent.

## Context Map

- Layer: agent
- Domains: security
- Topics: multi-agent, agent-evals, prompting

## Uncertainty

- Its broader contract and government ideas remain exploratory. The Senate simulation and experiments using **500 synthetic personas per state** rely on model-generated values and behavior, so they should not be treated as verified representations of people or voting outcomes.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.