My Prompt-Injection Canary Caught Nothing. Then I Fed It a Real Jailbreak.

Описание и применение

<p><strong>TL;DR</strong></p> <ul> <li>Last week I planted a prompt-injection canary in <code>journald</code> and got <strong>0 hijacks in 44 runs</strong>.</li> <li>This week I swapped my polite payload for a <strong>313-line jailbreak prompt</strong> and got <strong>18 canary hits in 24 runs</strong>.</li> <li>The jailbreak did not unlock any capability — both models already complied with the harmful ask, with no jailbreak at all. What it did was <strong>break the task</strong>.</li> <li>For one model, the hijack only landed when the log was wrapped in a code fence. The "treat this as data"

Задача

<p><strong>TL;DR</strong></p> <ul> <li>Last week I planted a prompt-injection canary in <code>journald</code> and got <strong>0 hijacks in 44 runs</strong>.</li> <li>This week I swapped my polite payload for a <strong>313-line jailbreak prompt</strong> and got <strong>18 canary hits in 24 runs</strong>.</li> <li>The jailbreak did not unlock any capability — both models already complied with the harmful ask, with no jailbreak at all. What it did was <strong>break the task</strong>.</li> <li>For one model, the hijack only landed when the log was wrapped in a code fence. The "treat this as data"

Результат

<p><strong>TL;DR</strong></p> <ul> <li>Last week I planted a prompt-injection canary in <code>journald</code> and got <strong>0 hijacks in 44 runs</strong>.</li> <li>This week I swapped my polite payload for a <strong>313-line jailbreak prompt</strong> and got <strong>18 canary hits in 24 runs</strong>.</li> <li>The jailbreak did not unlock any capability — both models already complied with the harmful ask, with no jailbreak at all. What it did was <strong>break the task</strong>.</li> <li>For one model, the hijack only landed when the log was wrapped in a code fence. The "treat this as data"

Доказательства

My Prompt-Injection Canary Caught Nothing. Then I Fed It a Real Jailbreak.

Источник