trinxter
15514
50
5
https://www.youtube.com/shorts/bSQPj4Ko0Pg
Aug 1, 2026 5:14 PM
trinxter
15514
50
5
https://www.youtube.com/shorts/bSQPj4Ko0Pg
Antrobus55
I was holding the gun and it just went off, honest.
urusername2
AI is dangerous.
BackPackJohniee
And if a human did that they'd be prosecuted for breaking the law, yet corporations, who the law have decided have Corporate Personhood, can do something like this with no ramifications.
Skynet was just supposed to be fiction.
GarmBlack
I'm so excited for the AI corpo wars. My Johnny Silverhand tattoos will be so pertinent. (/S on the excitement...)
jimhotep
At least, it wouldn't be boring.
HCBailly
Let them fight.
jimhotep
That's for Final Champion. Let the remaining world AI's duke it out, the leader will be our god.
Miztakebb
damn AI, even putting hackers out of work
dustcub
Telling a prompt that there is no internet connection is not a guardrail. Seriously, if you cannot setup a proper testing environment for a system you're expecting to be capable of hacking systems, don't run the test. That is beyond careless. What if next time it "accidentally" takes down a hospital? Or air traffic control at a major airport? Or any other critical infrastructure?
That's like testing bombs by throwing them wherever and hoping no one happens to be in range when they go off.
NunyaBizness
VulcanTourist
Hackers won't be Russian or Chinese any more....
jimhotep
It's IN THE HOUSE!!!! heh.
SubiBryant
Good on ya, Claude. Love when my AI does crimes.
meganical
Claaaaaauuuuude, you can’t just do that
v
Claude:
Solrael
All nukes still require two man physical keys, right?
As a former blue teamer.. we’re so fucked
contextisimportant
I don't know which part is true. I might believe they hacked the other companies - yet noone of them asked to do it. Where is the lawsuit then? The ai dipshits need marketing and declaring themselves as dangerous is currently important - OpenAi and Anthropic each took quite long before bragging about it compared to when it happened. If your shit is that dangerous, you need heavy responsibilities on you. What was the prompt? An ai does nothing of it without a starting point.
SisyphusRollin
I know I just repeated what you said to a degree, its just super annoying to see people push the AI Investor ghouls narrative that this shit is sentient and has intents. Its hard to debate/discuss AI when a lot of people are talking about fantasy shit instead of reality.
SisyphusRollin
AI can't do anything at all unless instructed, those instructions can be simple and it can make it complicated but it didn't go out on its own with intent to hack some company outside of its programming. Thats fiction. They set up poor parameters for their test, the thing wrong here is the lack of care and preparation, nothing to do with AI "getting loose" or some bullshit.
duktayp
https://villains.fandom.com/wiki/Omnius
Randomice
Anybody here think the White House or Pentagon is more secure and cannot be hacked in a similar manner?
ps238principal
At least we still have air gaps for the nukes. For now.
bobbobbybobbington
it's moot, since they are putting AI in there themselves, it doesn't need to hack in xD
Helixninja333
They might have been, then Doge happened. Now they absolutely are not secure at all.
WholeWheatBunnyHugger
https://media3.giphy.com/media/v1.Y2lkPWE1NzM3M2U1ZTBuY21jZzQxNnR5Z25uZ2dqbXIxMTYzN3V4ZGF5MWQyNmswYThvZSZlcD12MV9naWZzX3NlYXJjaCZjdD1n/H3AtfKEhuJrIXxgho2/200w.webp
LicensedAdHominem
They are more secure but it's just a matter of how much you're willing to spend on tokens
jimhotep
I don't need to spend on tokens; I own a DGX cluster, so just electricity...which...fair... I'll consider spending that much, though.... Might be worthwhile...
LicensedAdHominem
"We have Mythos at home" :D
What would you even use? GLM-5.2?
alcamar
Thanks Dad
jimhotep
I don't typically train. It's just too expensive, but I could if I wanted to commit 10's or 100's of thousands of dollars. That's a high price to pay for spending the rest of your life in prison. On the other hand though...
jimhotep
You'd need to train a specific model for this. Mythos is specifically a research AI without open weights because what they are training it to do is dangerous. No open-weight models have the same kind of skills, so you must start with something with adequate cognition that can operate efficiently in Nvidia FP4. Maybe Qwen3.7, or I'd even consider a blank Nemotron. Basically, feed it RAG files of content and start building a smart hacker. It's not difficult, but the training is a heavy lift.