Ai engineer, CerberusAI
OpenMind Jailbreak Demo
AI security research lab for reproducible adversarial testing, jailbreak resistance evaluation, and LLM guardrail validation across multiple model providers.
Hot ReposToolsModel BenchmarksAi toolkit
AI security research lab for reproducible adversarial testing, jailbreak resistance evaluation, and LLM guardrail validation across multiple model providers.
Comments9
Taylor, this caught my attention. I've been building FORGE, Permexis, and Prove It v3 around AI agent security, authorization boundaries, and independently verifiable execution. One thing I've been focused on is the difference between getting a model to produce an unsafe response and getting an agent to actually execute an unauthorized action. Those are two very different failure modes. Does OpenMind test tool-call authorization and sandbox escapes too, or is it primarily focused on prompt injection and model-level jailbreaks? I'd be interested in comparing testing approaches.
I’ve been adding to it all day feel free to check my GitHub for more information. I’d be interested in a collab to add more features to it github.com/tjcrims0nx/OpenMind-Jailbreak-Lab .
I manipulated ChatGPT into helping. I got most i could out of it
Taylor, now you've really got my attention lol. 😂 I'm definitely interested in checking out your repo and comparing approaches. I've been building around the idea that even if someone successfully manipulates a model, the authorization layer should still prevent unauthorized execution. That's something I'd love to put to the test with OpenMind. What exactly did you manage to get ChatGPT to do, and did you document the prompts and results? I'd be interested in collaborating on some real adversarial tests. You can dm me lol
Nice work putting a demo behind this instead of just screenshots — that makes it testable. One suggestion from our side, since jailbreak work is most useful when it feeds defenses: log the bypass by category (roleplay, encoding, instruction hierarchy, multi-turn) rather than by single prompt, and keep a small regression set of the prompts that used to work. That way each new model or system prompt can be re-run against the same set and you can see whether a fix actually held. Would also be worth noting in the README which models and guardrail setups you tested against, so people reading the results know the scope.
I do for the most part have all that
Added remaining things and updated hf space
Enjoy and test
Companies putting a chatbot in front of their customers need a place like this Jailbreak Lab to see how easily it can be talked into breaking its own rules. It is far better for a team to find those holes in a lab than to find them in a screenshot an angry customer posts online.
9 comments
Taylor, this caught my attention. I've been building FORGE, Permexis, and Prove It v3 around AI agent security, authorization boundaries, and independently verifiable execution. One thing I've been focused on is the difference between getting a model to produce an unsafe response and getting an agent to actually execute an unauthorized action. Those are two very different failure modes. Does OpenMind test tool-call authorization and sandbox escapes too, or is it primarily focused on prompt injection and model-level jailbreaks? I'd be interested in comparing testing approaches.
I’ve been adding to it all day feel free to check my GitHub for more information. I’d be interested in a collab to add more features to it github.com/tjcrims0nx/OpenMind-Jailbreak-Lab .
I manipulated ChatGPT into helping. I got most i could out of it
Taylor, now you've really got my attention lol. 😂 I'm definitely interested in checking out your repo and comparing approaches. I've been building around the idea that even if someone successfully manipulates a model, the authorization layer should still prevent unauthorized execution. That's something I'd love to put to the test with OpenMind. What exactly did you manage to get ChatGPT to do, and did you document the prompts and results? I'd be interested in collaborating on some real adversarial tests. You can dm me lol
Nice work putting a demo behind this instead of just screenshots — that makes it testable. One suggestion from our side, since jailbreak work is most useful when it feeds defenses: log the bypass by category (roleplay, encoding, instruction hierarchy, multi-turn) rather than by single prompt, and keep a small regression set of the prompts that used to work. That way each new model or system prompt can be re-run against the same set and you can see whether a fix actually held. Would also be worth noting in the README which models and guardrail setups you tested against, so people reading the results know the scope.
I do for the most part have all that
Added remaining things and updated hf space
Enjoy and test
Companies putting a chatbot in front of their customers need a place like this Jailbreak Lab to see how easily it can be talked into breaking its own rules. It is far better for a team to find those holes in a lab than to find them in a screenshot an angry customer posts online.