We're building a community library of agent attack scenarios, mapped to OWASP.

We're building a community library of agent attack scenarios, mapped to OWASP

Every agent has a failure mode somewhere. Maybe it takes a tool call it shouldn't, follows an instruction buried in a document it was only supposed to summarize, or holds onto something in memory it should have forgotten. Most of these get found once, fixed quietly, and never written down anywhere another builder could learn from.

We want to change that at Humanbound. We're starting a community library of "test packs": real scenarios where an agent gets pushed off-task or exploited, each mapped to a category from the OWASP Top 10 for Agentic Applications, and ideally paired with the guardrail rule that closes the gap.

Why OWASP mapping matters

A scenario on its own is useful. A scenario tagged to a recognized risk category (Agent Goal Hijack, Tool Misuse, Memory Poisoning, and so on) is something the whole community can build a shared vocabulary around, and something teams can use to check their own coverage against a known list instead of guessing.

What we're asking for

We've got a rough starter template rather than a rigid spec, since we'd rather refine the format from what people actually submit than lock it down up front. If you've found a hole in an agent you've built or tested, head to our GitHub Discussion and share:

• Scenario. What the agent does, and how it gets pushed off-task or misused. Enough detail that someone else could reproduce it.
  OWASP category. Which category it maps to. Not sure? Take your best guess, we'll help place it.

• Guardrail (optional). A rule or check that closes the hole, if you have one. Not required, a well-described scenario is a contribution on its own.

• Framework/agent type. What you were testing against.

https://github.com/humanbound/humanbound/discussions/125

This isn't a contest and there's no leaderboard. It's the start of a reference library other builders can point to. Everyone who contributes gets credited, and we'll use what comes in to shape a proper template over time.

If you've broken an agent in an interesting way, we'd like to hear about it.