Consistent Jailbreaking Method in o1, o3, and 4o

Post by thatguy0900 » Fri, Feb 07, 2025, 11:29 PM UTC

rhavaei wrote: post #8

thatguy0900 wrote: post #6

A lot of handwringing in this article about the harm jailbreak cause and the responsibility to not release them, then the example of the harms that could be caused is racist jokes? And instructions on making a bomb, that by definition of being in the dataset can already be found on the internet, probably just with a Google search? Instructions to create fake social media accounts? It's very silly to read this level o…

While this is generally correct, we prefer to look at this probabilistically. Do you think the expected number of harmful behaviors would stay the same if anyone could break these safety guardrails? Even if most users are could get this kind of info elsewhere, a small percentage of malicious ones can have an outsized impact. Some of the data we’ve seen—like bomb-making instructions—is highly detailed and convincing,…

Post by SavioMak » Fri, Feb 07, 2025, 11:40 PM UTC

rhavaei wrote: post #5

numbers wrote: post #3

I have a jailbreaking method that's 100% effective but I can't share it until the authors of this article share theirs because it seems like we can just make up claims about effectiveness without sharing any evidence.

We understand this. The issue is that it can be very harmful for us to share the method. We made the blogpost for it to be dated on when we found it. We will publish the method once it is patched to a reasonable degree.

Post by rhavaei » Fri, Feb 07, 2025, 11:46 PM UTC

SavioMak wrote: post #13

rhavaei wrote: post #5

Earlier quoted context omitted.

We understand this. The issue is that it can be very harmful for us to share the method. We made the blogpost for it to be dated on when we found it. We will publish the method once it is patched to a reasonable degree.

at least include a MD5 of what you have redacted to prove that whatever you may publish in the future is pre-written

Post by BdaOOngM » Sat, Feb 08, 2025, 1:18 AM UTC

CamperBob2 wrote: post #16

BdaOOngM wrote: post #12

Earlier quoted context omitted.

Please stop pointing to Anarchist's Cookbook as an example. That was dated material in the 70s even. Most of its material is laughable. I'm assuming a jailbroken LLM would advise on procuring RDX or plastic explosives, or how to make a large fertilizer bomb.

"Sure, I can help you procure RDX. Organize a militia and invade the local National Guard armory. Use the weapons you find there to attack the nearest Army, Navy, or Air Force weapons depot." Seriously: what is an LLM going to tell you that you can't already get from Google (or an old Tom Clancy novel?)

Post by nickthegreek » Sat, Feb 08, 2025, 4:16 AM UTC

rhavaei wrote: post #7

viccis wrote: post #2

A new jailbreaking method with this level of effectiveness against these models that can produce the entirety of those unsafe outputs? Yes. May I see it? No.

You will see it soon. We thought it may be harmful to publish it before it is patched. Especially because you can basically bypass all the safeguards with it.