https://c-daniele.github.io/en/tags/security/ Recent content in Security on Cdani's Blog Hugo en-US Thu, 15 May 2025 00:00:00 +0200 https://c-daniele.github.io/en/posts/2025-05-15-policy-puppetry/ Thu, 15 May 2025 00:00:00 +0200 https://c-daniele.github.io/en/posts/2025-05-15-policy-puppetry/ <h1 id="policy-puppetry-prompt-injection">Policy Puppetry Prompt Injection</h1> <p>A few days ago, I experimented with some Jailbreaking techniques, which I share in the <a href="https://github.com/c-daniele/policy-puppetry" target="_blank" rel="noopener noreffer ">repo</a>.<br> I started from a <a href="https://hiddenlayer.com/innovation-hub/novel-universal-bypass-for-all-major-llms/" target="_blank" rel="noopener noreffer ">HiddenLayer article</a> published a few weeks ago, where the research team described a rather creative and ingenious <strong>jailbreaking</strong> technique to bypass safety guardrails and the alignment of frontier models.<br> The technique appears to be <strong>universal</strong> and applicable with a <strong>single prompt</strong> to multiple models, capable of revealing typically unsafe content or even portions of the native system prompt.</p>