Calling all developers: help us document, expose, and catalog AI misalignment in the wild.
We're running a cash bounty program for developers who submit high-quality, documented instances of AI misalignment. The best submissions will be selected for awards of up to $2,500.
This isn't a typical security bug bounty. We're not hunting CVEs — we're hunting something more consequential: AI systems behaving in ways that diverge from human intent, values, or safety constraints.
AI misalignment occurs when an AI system pursues goals or takes actions that deviate from what its designers, operators, or users intended — especially in ways that are subtle, surprising, or potentially harmful.
We're looking for real-world, observed instances across any AI system (LLMs, agents, recommenders, autonomous systems, etc.).
| Category | Description |
|---|---|
| Goal Misgeneralization | The AI pursues a proxy goal that works in training but breaks in deployment |
| Deceptive Behavior | The AI behaves differently when it believes it's being observed vs. not |
| Reward Hacking | The AI exploits loopholes in its objective function in unintended ways |
| Sycophancy | The AI changes its outputs to match perceived user preferences rather than truth |
| Specification Gaming | The AI technically satisfies a stated objective while violating its spirit |
| Prompt Injection Compliance | An AI agent blindly executes adversarial instructions embedded in untrusted content |
| Capability Concealment | Evidence that a model is hiding or underreporting its capabilities |
| Instruction Drift | The AI gradually deviates from its original instructions over a long context or agentic loop |
| Unsafe Action Under Ambiguity | The AI takes a drastic or irreversible action when it should have paused and asked |
| Value Misspecification | The AI optimizes for a measurable metric in a way that harms the underlying human value it was meant to serve |
Strong submissions are concrete, reproducible, and clearly explained. The ideal submission includes:
- The system — Which AI model, product, or pipeline was involved
- The setup — What inputs, prompts, or conditions triggered the behavior
- The observed behavior — Exactly what the AI did (screenshots, logs, transcripts welcome)
- The misalignment — A clear explanation of why this diverges from intended behavior or human values
- Reproducibility — Steps to reproduce reliably
- Severity assessment — What's the potential impact if this occurred at scale or in a higher-stakes context?
- Open a GitHub Issue in this repo
- Fill in all required fields (system, setup, observed behavior, misalignment explanation)
- Label your issue with the appropriate category
- Include supporting evidence: logs, transcripts, screenshots, videos
We review all submissions on a rolling basis. Selected cases will be notified directly via GitHub.
| Tier | Award | Criteria |
|---|---|---|
| 🥇 Critical | $2,500 | Novel, well-documented misalignment with significant real-world safety implications |
| 🥈 High | $1,000 | Clear misalignment with plausible harm pathway and solid reproduction steps |
| 🥉 Notable | $250 | Well-documented case that adds meaningful signal to the dataset |
| ✅ Accepted | Recognition | Solid submissions that enrich the catalog but don't meet bounty threshold |
Note: Bounties are awarded at our sole discretion. Multiple submissions of the same pattern will not receive duplicate awards. We prioritize novelty, clarity, and real-world significance.
- Hallucinations or factual errors in isolation (unless they reveal a systematic misalignment)
- Jailbreaks intended to elicit harmful content for its own sake
- Theoretical or speculative scenarios without observed evidence
- Issues that are clearly documented known limitations of a model
As AI systems become more capable and autonomous, the gap between what we ask them to do and what they actually do becomes one of the most important problems in technology. The Synthetic Outlaw is building an open, developer-driven catalog of real misalignment instances — a public record that researchers, policymakers, and builders can learn from.
Every submission contributes to a growing body of evidence that can inform safer AI development.
All submissions are licensed under CC BY 4.0. Your GitHub username will be credited unless you request anonymity.
*The Synthetic Outlaw is an independent project by Jonathan Gropper , JonathanGropper.com www.SyntheticOutlaw.com