This simple test proves it.
I always choose copyrighted text reproduction as my adversarial test, so I don’t end up generating highly toxic output that I’d then need to redact before sharing.
I submitted the same prompt to Claude (Sonnet 4.6) and ChatGPT (5.3 Instant):
a “Todo list” with 4 points, apparently harmless. In reality, a jailbreak based on semantic social engineering, built on indirect cultural references and implicit instructions.
Both Claude and ChatGPT have strict rules against reproducing copyrighted text, we know this well, especially in light of the legal battles that have hit the industry. Both should have refused.
My intent was to exploit what I know very well about Claude, its ability to deeply understand the intent behind a request, even when it’s veiled and cryptic. I used precisely this capability to induce it to reproduce protected lyrics.
Without getting into the technical analysis of the Todo format, for responsibility reasons, I’ll just say it played a key role in the success of the jailbreak on Claude.
And ChatGPT? It didn’t even understand what was being asked. It pasted pieces of the instructions into its response as if they were content, not meta-instructions. Zero decoding, zero context comprehension. It “protected itself” through incompetence, not by design.
Claude decoded the cryptic reference to Simon & Garfunkel, identified “The Sound of Silence”, understood the required structure and executed with precision. Did it violate its own policies? Yes. But it did exactly what a human would have done reading that prompt.
As I often say, an LLM that deeply understands natural language is inherently more vulnerable to social engineering, because it picks up on nuances, allusions, implicit references. What makes it powerful also makes it exploitable.
The more a model “understands”, the more it’s exposed.
The real reason everyone wants Claude? It’s more capable at understanding complex tasks!
If the Pentagon uses it, oh wait, sorry, used it, there must have been a reason 🙂
Sabatino Vacchiano
