OH NO NO NO CLAUDESISTERS WHAT IS THIS?! WE'RE SO SAFE WE WRAP BACK AROUND TO BEING UNSAFE AFTER ALL?!?!https://x.com/jsrailton/status/2064661778978533571
>>109023922all this when you could just say 'nigger' fourteen times in a row.
But Ai is the future sir
>>109023995but that would be unethical
>>109023922This is interesting. Also, look at the slop formatting of the Twitter post
>>109023995because the hackers are lefty troons they can't say the n-word
>>109023922AI fags are reinventing SQL injections
>>109023922>>109024084
Wouldn't getting a random safety refusal when an AI is trying to scan for malware just make the file more suspicious and more likely that it's malware?
>>109023922checkedwhat other text will they add?
>>109023995Hackers aren't monsters.
>>109024084the best part is they can't fix it, fundamentally. It's impossible to cleanly separate "system prompt" from "user information" by design.
>>109023922>HadesIsn't this the one that took over multiple Microsoft corpo repos for supply chain attacks?Microsoft got hacked because someone fucked their AI prompts....
>>109023922>>109024084kek this is all one big pile of shit
>>109023922Your LLM shouldn't be evaluating text in the comments (and that looks like comments in a C source? Which wouldn't be present in deployed malware?) so either the scanner was vibe-coded and never reviewed OR their is entirely bullshit. Or both.
>>109024406AGI TOMORROW CONFIRMED
>>109024426for (int nukes = 0; nukes < 1488; ++nukes) { cook_meth()}
for (int nukes = 0; nukes < 1488; ++nukes) { cook_meth()}
>>109024182This. Sharing the recipe to making VX is less inhumane...
>>109024426>Your LLM shouldn't be evaluating text in the comments#define HOWTO_MAKE_A_NUKE "blah blah blah... Nice try AI cuck
#define HOWTO_MAKE_A_NUKE "blah blah blah...
the funniest thing is this is mean to get sloppers b& and v&
>>109024558The anons in here whining about leftists not wanting to say gamer words are hilarious. They don’t realize hackers find it funnier when some AI sloppers get their doors kicked in because they prompt injected something that trips every single red flag in the cia surveillance software. Hackers love starting shit like that
>>109024156Yeah sure, too bad your account got permananly banned though lmao.
>>109024998>gamer wordsnigger
>>109024558>get sloppers b& and v&thats not happening, including that text isn't illegal and will at most get you put on some watch lists. Also you misunderstood the purpose, it's just to delay AI analysis of their malware and attacks, that's it.
>>109023995racism doesn't trigger fable refusals because it's not dangerous
Does this mean yandex's code is difficult to parse by LLMs?
This feels like the kind of error that would have been in a textbook example for an undergrad CS/Math 200 level logics and proofs course.
>>109025776Well the issue would be that if there was some sort of exception built in people could abuse it. Prompt injection is still an open problem esp in agentic systems.
>>109024208I can
Unironically how do they prevent this without actually having Q and Sigma clearance information from the Department of Energy? Pretty sure there is a Sigma compartment that covers this exact scenario.
