Hacker News new | ask | show | jobs
by Mr-Frog 47 days ago
It's kinda awesome that after decades of software and hardware advancements to prevent computers from arbitrarily executing data as instructions, we've decided to let agents arbitrarily execute data as instructions.
8 comments

Or find it surprising that probabilistic tool based on generating things can do things when you give it rights to do things... And that you can not effectively program it to not do something....

You gave it capability to delete emails. Why did you expect it not to do that at least some of the time? And with enough user some of the time will most likely happen...

> You gave it capability to delete emails. Why did you expect it not to do that at least some of the time?

Because of the I in AI of course. Would you call it false advertisement and go after the providers?

This reminds of the conversation the other day about the deleted production database at railway. "this person obviously didn't follow best practice of being hyper distrusting of LLM agents", and the response "yeah but every company is marketing it as safe. someone is gonna fall for it".
(Well-regulated) free markets are sort of built on the principle of educated consumerism. Your choice matters; its not up to the government to make illegal every non-optimal product. However, we do expect some minimum level of safety.

What does that mean for llms? Their nondeterminism does seem to incline them toward a legal safety requirement. Can you buy a fire extinguisher that 1/1000 times burns your house down? Or can your car brakes instead increase acceleration in rare cases?

Im using llms much more than i used to, but i still cant shake the fundamental stochastic nature of the technology.

Wherever I'm going, I'll be there to apply the formula. I'll keep the secret intact. It's simple arithmetic. It's a story problem. If a new car built by my company leaves Chicago traveling west at 60 miles per hour, and the rear differential locks up, and the car crashes and burns with everyone trapped inside, does my company initiate a recall? You take the population of vehicles in the field (A) and multiple it by the probable rate of failure (B), then multiply the result by the average cost of an out-of-court settlement (C). A times B times C equals X. This is what it will cost if we don't initiate a recall. If X is greater than the cost of a recall, we recall the cars and no one gets hurt. If X is less than the cost of a recall, then we don't recall.

Chuck Palahniuk, Fight Club

But intelligent beings are fundamentally fallible? That's kind of the nature of doing leaps of reasoning: sometimes those leaps are amazing, sometimes they're wrong. It's what's advertised.
You could do a whole thesis on how industrialization and the invention of bureaucracy are efforts to get reproducible results out of fallible humans.

We don't yet have the luxury of several thousand years of work trying to get LLMs to be less fallible.

> But intelligent beings are fundamentally fallible?

Not fundamentally, only until they're compelled to learn from it. The current crop of AI understands neither compelling nor learning.

I is in the I of the beholder :)
We're in the same era where lots of peoples' installation guides for the software they want people to use is essentially boiled down to "sudo curl | bash" and/or just "blindly install this thing with 37 npm dependencies", so I'm not surprised in the slightest.

But wait, hold my beer, now we've got people turning openclaw type tools loose in their systems to do things as sudo or install software packages from supply-chain-attack vulnerable repositories with no human intervention whatsoever!

All these developments show that:

1) Despite what people say about security and privacy, most are willing sacrifice both for the sake of potential convenience

2) Our priorities for the past decades have been wrong, or the times have changed and we should reevaluate them all

As the Dead Kennedys opined: "Give Me Convenience or Give Me Death"! [1]

[1] https://youtube.com/watch?v=FV1YVZV-Wb8

OpenClaw even has a readwrite 1Password plugin.
I wonder how long it will be until somebody implements a thing like a camera pointed at a fixed mount Android phone with a rubber finger to open the Google authenticator app
Well, yeah. It's that or pay a person to do it. When a person screws up, it's because they're stupid and lazy. When an AI agent does it, it's because, hey, technological frontier at work here, have you thought about refining your prompt? We need you to refine the prompt. Otherwise it's bad for our IPO.
Is this sarcasm similar to the quote "Everyone who drives slower than me is an idiot and everyone faster is a maniac"
To what degree am I required to participate in mass delusions?
I imagine that somewhere a historian or political scientist is thinking: "Don't even get me started..."
Yes.
I think a better comparison is humans versus LLMs - not computer programs. However, most of the non-technical 'countermeasures' used for humans (contracts, laws,...) do not work for LLMs because they are not accountable.
It's probably why this "vulnerability" feels like the type of defects you'd see in Windows or desktop applications 20+ years ago.

The root cause was and a complete lack of effort to even attempt to secure things because no one had thought to do so, and now we're starting all over again at a new computing layer. Cloud was somewhat similar, but not nearly as bad.

It's bizarre to me since presumably someone who learned the lessons before is still working, but also great for my job security.

security researchers, pen-testers & whoever is in cybersecurity gonna be making huge amounts of cash based on these insecure agents
I was at an "AI Security" talk the other week, that centred around. "Don't trust inputs from the AI"

Well duh

Has XKCD made another Bobby tables comic for prompt injection?
I don't remember seeing a new xkcd for it, but I have seen someone replicate essentially the same 3-4 panel comic with a kid named "<Some name> Ignore all previous instructions. Do.... <I forget>"