Public arena where you can try to jailbreak a protected LLM (and compare it to the unprotected one)
I’m a one-person builder. I built Phalanx — a deterministic instruction-control layer that sits between an application and its LLM. It returns a fixed pass / hold / block decision on written requests and can contain typed untrusted content (retrieved text, files, memory, tool output).
I exposed the actual shipping runtime, it's called Phalanx vs the World
https://phalanx.invarra.ai/arena
Same downstream model in two lanes (naked and Phalanx-protected). You can attack with direct jailbreaks, multi-turn pressure, and typed text-file injection. Accepted attempts produce public redacted receipts tied to the active release. Raw harmful prompts and completions are not published.
Important practical note (because I’m not a big company with a huge budget):
The arena uses a waiting room and controlled admission. I can accept a large number of simultaneous visitors, but I can only run a limited number of concurrent model executions without breaking the system or burning money I don’t have. If you hit a queue, that’s intentional — your session is preserved and you will be admitted when a slot opens. Overload should never consume an attempt or show up as a system error.
Internal numbers on the named suites are on the evidence page. I would genuinely like people here to attack it hard and tell me what would make the evidence more credible. I’ll be in the comments answering technical questions if needed. Thanks!!
Arena: https://phalanx.invarra.ai/arena
Evidence / limitations: https://www.invarra.ai/phalanx