AegisLM
See how easily your AI can be broken — in seconds
5 followers
See how easily your AI can be broken — in seconds
5 followers
AegisLM shows how easily modern AI can be broken. Test any model for prompt injection, jailbreaks, and data leaks in seconds. Input a prompt, run attacks, and see where it fails. Designed for builders who want to stress-test AI systems under real-world conditions. Try built-in attacks or create your own.
RiteKit Company Logo API
@aca_050 This is a genuinely important angle—most teams obsess over capabilities and miss the failure modes. The adversarial testing approach makes sense, especially since prompt injection and jailbreaks are real production risks now. For the next phase, I'd suggest testing against models with different safety training, since vulnerabilities often vary by architecture and training approach.
@osakasaul Yeah, I've started observing that too. The behavior of the models varies greatly when they are given the same prompt based on their alignment or the structure of the system prompt.
Right now I’m mostly focused on making the failure cases easier to reproduce consistently, but comparing models with different safety training approaches is definitely something I want to dig into more.
Specifically, I would love to see how an open model behaves relative to a heavily aligned API model under the same prompts.