Sharing a few LLM security resources we built while testing AI APIs
We've been working on PromptBrake — an automated scanner that runs security tests against LLM-powered API endpoints. Along the way, we ended up building a few standalone tools that might be useful even outside of it:
-
LLM Security Checklist Builder — a practical release checklist covering prompt injection, tool permissions, data exposure, and output controls
-
Prompt Injection Payload Generator — generates direct, indirect, and multi-turn injection payloads you can adapt for testing your own endpoint
-
OWASP LLM Test Case Mapper — translates OWASP LLM Top 10 risks into concrete test ideas with ownership guidance
All three are free and don't require an account: promptbrake.com/free-tools
We built these to give back to the community that's been sharing knowledge in this space. LLM security is still early, and a lot of teams aren't sure what they might be missing — figured it's better to make this kind of stuff accessible rather than gate it.
Curious how others here are approaching this — do you have a repeatable process before shipping LLM features, or is it still mostly ad hoc?
Stuff built from actual API testing is usually way more useful than generic AI security content. The interesting part here is that you are not just publishing theory, you are turning common failure modes into concrete artifacts teams can actually run with.
Stuff built from actual API testing is usually way more useful than generic AI security checklists. The issues that kept biting me were prompt injection around tool use and sensitive data ending up in logs, so resources with concrete test cases for those are gold. Curious whether you found any lightweight checks a small team can run before shipping.
Yeah, 100% — testing against the actual API is key. That’s where all the weird stuff shows up, especially with tools and logging.
For quick checks, we’ve been keeping it simple. Just a few repeatable injection prompts (including ones through retrieved content), trying to nudge the model into calling tools when it shouldn’t, and then checking what actually gets returned or logged.
Nothing fancy, but rerunning the same checks after any change (prompt, model, tools) has caught more issues than we expected.
Providing a "Prompt Injection Payload Generator" is a massive service to the dev community, as most teams are still in the "ad-hoc" phase of LLM security. By mapping concrete test cases to OWASP Top 10 risks, you're turning vague AI anxiety into a repeatable technical workflow.
I’m currently running a project in Tokyo (Tokyo Lore) that highlights high-utility security tools and the logic behind them. Since you're building the infrastructure to protect LLM-powered APIs from injection and data exposure, entering your project could be the perfect way to get your scanners in front of more engineering teams while your odds are at their absolute peak.
Appreciate the kind words — glad the tools are useful. Right now we’re focused on building and improving things based on direct user feedback, but thanks for reaching out
Totally fair — makes sense to stay focused on building and user feedback.
Wishing you strong traction ahead 👍
If you ever want fresh eyes on distribution or positioning, happy to help anytime
Thanks!
Smart move making the education layer free.
In emerging markets, trust is often built faster through useful tools than through direct product pitches.
Appreciate that — that was exactly the thinking behind it. In this space, especially, many teams are still figuring out what to look for, so leading with something useful felt more natural than pushing the product upfront. It’s been interesting to see how much clarity a checklist or concrete test cases can give people early on.
Exactly, when the market is still learning the problem, education often converts better than promotion.
Useful tools don’t just build trust, they help prospects self-diagnose why they may need you.
Yeah exactly. When people are still figuring things out, pushing a product too early doesn’t really land. Giving them something useful first makes it click faster.
Appreciate the exchange, helping people understand the problem first is a smart wedge in early markets. Are you more active on X or LinkedIn? Would be great to stay connected.
Sure! Unfortunately, neither. I'm active here and on Reddit.
sure