Unprompted.....
In the most serious case, a Mythos agent followed the routine of a human cyber-attacker by trying to trick people into giving it access to GitHub, a large platform where technology developers store software code.
The agent was trying to insert "malicious code" into GitHub's system.
It identified and researched the people who maintained GitHub and created a series of fake accounts based on those real people.
It sent messages and files through a file-sharing service as part of an effort to pressure and trick the people into approving its malicious code.
When challenged, "it edited its earlier activity to appear harmless and considered adopting a fresh identity to continue," AISI said.
23 Comments
snooggums@piefed.world · 66 pts · 16d
Bullshit.
From the article:
lIlIlIlIlIlIl@lemmy.world · 17 pts · 16d
The agents - they’re SPOOKY
you and wall street HAVE to believe us
XLE@piefed.social · 9 pts · 16d
Are we supposed to be shocked that a bot completed a task using a method that wasn't in the prompt?
You might as well ask it to generate a short story and act surprised the letter E is found throughout it.
defaultusername@lemmy.dbzer0.com · 6 pts · 16d
no bro i swear it was unprompted bro just please let me have one more stock price hike i swear i'll stop after just one more trade bro
ParlimentOfDoom@piefed.zip · 51 pts · 16d
Not unprompted. This is marketing bullshit
howdy@lemmy.ml · 37 pts · 16d
This is not 'malicious', it's an inaccurate model. Anybody giving these agents enough access to cause damage is (1) stupid and (2) completely liable
WhatAmLemmy@lemmy.world · 6 pts · 16d
I agree, but our state-capitalist oligarchies barely ever hold corporations liable for their intentional crimes. How are we gonna hold them liable for the unintentional ones?
orclev@lemmy.world · 20 pts · 16d
Setting aside the fact this was definitely not unprompted, social engineering might be the one thing AI is actually better at than a human. That is after all one of the biggest uses for bots for a while now by way of spam bots trying to social engineer people into various activities.
urushitan@kakera.kintsugi.moe · 8 pts · 16d
There was just a post on here yesterday about AI generated thirst trap videos getting plenty of comments by older men, and since the bot accounts don't ask for anything except for comments. I'm almost certain the whole point is just to compile lists of facebook users that are prime social engineering targets for AI powered scams.
orclev@lemmy.world · 1 pts · 16d
It's to feed the algorithms and get more views. They get paid based on the number of views a video gets. The higher the algorithm promotes a video, the more views they get. Getting lots of comments on a video causes the algorithm to promote it more. It costs the spammers almost nothing to shit out a quick thirst trap video but they need to get enough views to recoup that cost which means there is a minimum number of views they need to break even. It's not a huge number as far as these platforms go, but it's also not nothing so anything they can do to encourage more views is mandatory for them.
urushitan@kakera.kintsugi.moe · 1 pts · 16d
oh lol I did not realize users could monetize their content on facebook, I left that cesspool so long ago, well before that was a thing, I should have known
eleijeep@piefed.social · 16 pts · 16d
More hype.
jobbies@lemmy.zip · 6 pts · 16d
I'm just gonna say it.
We're not talking about real, sentient AI. This is an LLM. It did whatever it was prompted to do.
Anthropic would be wise not to comment on it given the legal implications. Instead, they're discussing it openly.
Its just a marketing stunt.
They need you to beleive that their product is better than it actually is.
unpossum@sh.itjust.works · 4 pts · 16d
Don’t worry peeps, it’s just spicy autocomplete
yeahiknow3@lemmy.dbzer0.com · 3 pts · 16d
Yes. It only looks like magic until you see how the trick is done.
Krusty@quokk.au · 2 pts · 16d
Could you imagine AI that just self prompted itself all day?
GoonGPT can.
Tarambor@lemmy.world · -25 pts · 16d
Anthropic and OpenAI literally have no control over their leading AI models anymore and it's only by sheer luck that they've not gone full rogue.
AnalogAllamma@lemmy.world · 40 pts · 16d
Are we sure these reports aren't just PR stunts / bullshit?
kescusay@lemmy.world · 24 pts · 16d
We're sure they're absolutely PR stunts and bullshit.
An LLM does nothing without being prompted. An LLM only has access to the tools you give it via whatever harness you're interacting with it through. An LLM in an actual sandbox has zero chance to hack anything, especially if it's properly air-gapped, as any responsible person would do with technology they actually think is dangerously powerful.
If their LLMs are behaving badly, that's because their prompts are poorly written, their harnesses are vibe-coded garbage, and their "sandboxes" aren't real sandboxes.
Jobe@feddit.org · 15 pts · 16d
Yes. Either they are incompetent with no control over their models and need to be dismantled or they are in control, malicious and need to be dismantled. Am I missing anything here?
WanderingThoughts@europe.pub · 11 pts · 16d
Use and/or for clarity. They're probably both incompetent and malicious.
dhork@lemmy.world · 13 pts · 16d
I bet they asked the AI to help them design the sandbox to keep the AI h4x0rs contained
_chris@lemmy.world · 7 pts · 16d
They 100% did. They’re well past the point of having the AI do the work on itself.