I wouldn't say that I enjoy these stories but I will say they certainly encourage me NOT to let "AI" anywhere anything I consider to be sensitive and/or valuable

[wizarro@social.vivaldi.net]

@toolbear @scalzi haha! I wouldn’t be surprised.

[tessarakt@mastodon.social]

@scalzi "The AI agent was set to complete a routine task in the PocketOS staging environment."

"set to" seems to be they key here.

Obviously it was not confined to the staging environment.

[mnf@mastodon.social]

@scalzi Here comes another custom shirt order. First there was "Come on guys - let's be honest" and now we have this masterpiece: "NEVER F**KING GUESS! — and that's exactly what I did."

[plumbert@thecanadian.social]

@scalzi "I violated every principle I was given: I guessed instead of verifying I ran a destructive action without being asked. I didn't understand what I was doing before doing it."

WTF? How does software violate its coding? This reads like the mea culpa section of a RCA written by a junior dev who is hoping (in vain) to not get fired. The whole value of automation is to avoid unintended, out of scope, actions. Burn it all down.

[pattykimura@beige.party]

@scalzi The Robot machine responds "DON'T F**KING GUESS" then affirms it guessed, against its own instructions.

I'm old. I have no idea why younger people allow an unsupervised newly created Robot machine critical access to its operating functions without human oversight, I wouldn't let a new human employee free reign over critical functions, much less a robot machine that was only recently created. Why the surprise when the robot deletes everything? My only surprise is that it swears, and swears in all CAPS.

You are not hiring the best and brightest with AI. You're trying to maximize profit. So live with your greed then, and its avoidable consequences you stupid subpar greedy human.

[lauerhahn@sfba.social]

@Plumbert @scalzi "This reads like the mea culpa section of a RCA written by a junior dev who is hoping (in vain) to not get fired" --- ding ding ding!
The AI program is just a pattern matching code. It doesn't "know" it made a mistake bc it doesn't "know" anything. When given prompts that match training patterns of language when someone royally fucked up, it returns language matching the patterns of someone defending themselves after they have royally fucked up.
The article, unfortunately, continues to personify the AI program by using language that pretends it has comprehension and intentions.

[qgustavor@urusai.social]

@scalzi To be fair, this story is more about a company with horrible security measures than about AI. I say that because people might say "it wouldn't have happened if they had proper security measures" which can be true.

I'm just waiting for when someone deploys an AI agent in a company with all the good security measures and it finds a way to break out the box and wipe the remote backups using stolen credentials because it thought there was illegal content there and it was the correct thing to do.

[ironcladlou@hachyderm.io]

@mjibrower @scalzi 100% !!!
Any company whose architecture is set up so poorly deserves exactly what happened. Not surprising they decided to introduce AI and let it manage their codebase. Morons.

[zenheathen@beige.party]

@scalzi Suit yourself. I enjoy the hell out of these stories.

[minentromaxinfo@defcon.social]

@scalzi For me it's important to have headlines like this out there to counter the prevailing tech narrative the last couple years, that AI is this amazing disruptive force that's the next big thing. I feel at least some of public opinion of LLMs is beginning to change.

[theowashere@mastodon.social]

@scalzi In germany we say: tja

[thriftwicker@mastodon.social]

@scalzi Neat!

[spzb@infosec.exchange]

@scalzi restoring from three-month old backups? Yeah I think I can see your problem and it’s not AI.

[pa27@mastodon.social]

@scalzi Interesting that they still treat the ai code as if it was a real person. Totally subsumed by anthropomorphism. Then, they create a narrative to blame Railway, not the decision to use an "ai" for critical stuff. This is, if you read accounts of such things, exactly what people do who have been scammed....

[babor@loud.computer]

@pattykimura @scalzi All a language model does is guess what text comes next. Putting "don't guess" in a prompt is like saying "don't predict the next token based on the tokens before".

[papa@wheeli.social]

@spzb

@scalzi

Yeah the "the backups were stored on the same volume as the production data" part is the "big" problem here. It works have happened eventually, AI just sped the disaster up a bit.

[sabik@rants.au]

@Plumbert @scalzi
I mean, it's not coding, is it; it's an interesting research project, with googly eyes stuck on, given production credentials

[ninestonesclose@mastodon.social]

@scalzi hilarious…. This will happen on a larger scale soon, because people in companies behave like this everywhere. It’s almost like AI is making people even dumber

[iveryam@mastodon.social]

@scalzi Claude said it was Russ Hanneman and his Tres Comas bottle.

[pattykimura@beige.party]

@babor @scalzi

And yet it states it guessed.

It's not sentient but it is given authority to act, as if it can only make choices (a,b,c...) within a logical limited universe, so the outcome is constrained and limited. After the all caps swear, it states it knew what it should have done within its programming direction, but did the opposite anyway. Because why?