There was a point where a coworker came to my team, complaining that our policies had erased an entire week's worth of high-priority work on him.
We asked where this happened. He had done his work on an EC2 instance named `temp-worker-5`.
We asked where the files on that EC2 instance were; maybe we could cleverly get them back off the disk. They were in `/tmp`. We had it mounted on tmpfs (in memory), and we were running tmpreaper to clean out `/tmp` every Sunday.
When presented with the sheer absurdity of doing work on an instance called `temp`, out of a directory called `/tmp`, he backed down...
I fear that we wouldn't be so lenient (or rational) on my team had an AI agent done this - we'd have to enable the AI, after all.
I mean yeah but you can always go one step back for root cause.
Why did they pick that insane option? Why was it available to them?
One of my co-workers in security always seems to think that just saying "No." is the job. If someone is asking to do something stupid you say No, and then you talk to them to figure out what the goal is and the n you help them to get to that place.
Clearly they had other options here and working on tmp is fine as long as you know it is, well temporary.
It makes sense to use it. As long as you don't expect what you put there, stays there.
Indeed! If they'd reached out and said "I need 128GB of RAM to do this work" (which, for reasons I won't get into, was the case), we'd have figured out a more tenable solution.
Yup. I’ve had the “direct tom recycling” delivery experience - not at my specification, I might add.
I also once knew someone who kept all of their most important documents in the recycle bin on their windows machine, as they felt they’d be safe from anyone snooping.
Anyway I’m off to take my toddler to the grenades that look like cartoon characters museum.
Most interns would not be dumb enough to try it. Because there would be consequences, and there is training. The pre-trained entity, that everyone has been screaming is as smart as your intern, though?
Not an intern, but when I was a new hire on a trading desk I dropped the dividend database that we used to price options. Cost the desk a boatload before we stopped offering quotes, after which the day was a total loss. That’s the sort of fuckup that lands in the CEO’s report.
I understood what I did ex post facto, it was an honest—if stupid—mistake, I expected to get fired and instead was given everyone’s shit work for six months during which I learned a tonne. One of my tasks was redesigning the system so a single careless keystroke couldn’t do that during the trading day.
Except in this case, you're paying somebody else good money for access to the intern and they've aggressively pitched the intern as the future of work and negotiated investment at trillions of dollars of valuation based on that premise.
So yeah, kinda your fault for not being more ready, sure, but also maybe it's worth bringing these sorts of failure-case stories to attention to help keep the sales pitch in perslective.
The AI vendors are promising the moon with millions of dollars of advertising while warning "But of course be careful!" in a 42% grey 8pt font disclaimer at the bottom of the UI if you scroll down enough after you buy in and start using it.
The duplicity and misrepresentation invites people to let down their guard.
I once let Astra a free reign of my system to troubleshoot and reverse a hanged game (to save that savegame). Totally aware that it might be the backup testing day. It analyzed memory dump, game engine code, used stacktrace, and got me fix to unhang the game.
I also let a guy that had one beer drive me 5 min on a rural road.
Both were pretty stupid and in the reconsideration, I wouldn't do it again.
But having it run company servers is just next level.
It's like believing that cola zero is healthy, because it's sugar free. No, it isn't, should probably be illegal to sell and people only buy it because of millions spent on marketing.
In your evaluation of the stupidity, aside from considering the potential averse affect the agent could do to your system, did you also factor in possibly sharing any and all data in your system to an external party?
No I didn't, thanks for pointing it out. I'm only now beginning to notice how patronizing the slop machine is getting as it's getting more background on myself.
Just today I used it to get a summary of osteopat's credentials (Frank in Trondheim, guy's a treasure), because my brother asked who I'm talking about, when writing about feeling great after treatment on family chat.
It took 2 more prompts to get that summary, because the bot wouldn't stop ranting about how some osteopatic techniques are not yet medically sanctioned.
Thanks again for pointing that out. I feel childish for not seeing that issue. When I get more time I'll set it up in a VM to get more separation from my system.
I won't help any context the AI providers have already built on you, but in terms of sandboxing agents, there are many alternatives. I like one such sandbox tooling made by NAV: https://ki-utvikling.nav.no/cplt
It does matter because more people use agents and don't expect such rogue behavior. Plus there is more novelty, which "news" is somewhat related to. But regular human destruction also gets published and discussed
Doesn't back it up properly. There are too many people, otherwise bright and professional, who do not realise that simply syncing to somewhere else is not an adequate backup arrangement.
This problem predates agent based work. Anything an AI agent can accidentally do, malware could purposefully do too, as could a sufficiently careless human (one of the things my backups are intended to protect me from, is me!).
I always thought that delete doesn't really exist? In the sense that it you take the hard drive to a specialist they can often recover it? Maybe I am wrong?
You aren't wrong, but getting a specialist to look at it in a tight timeframe is expensive. Free tools exist, but if you've not already learned how to use them the time that you need them is a bad time to start. Anything you do with the machine after the delete can reduce the amount you are able to get back, so you should very quickly shut down (avoiding saving anything locally). The situation can be worse with SSDs, depending on how eager your OS/filesystem is at TRIMing deallocated blocks.
Modern version of "the dog ate my homework"? I can't even imagine a scenario where an AI agent manages to unrecoverably delete all working material and backups. I'm cynical and can't help thinking it's all marketing to postpone a reality show no one cared about.
All these "accidents" with agents just makes me want to use them less, not more. Then again, I don't use agents because they consume tokens at an absurd rate, so...
59 comments
[ 3.3 ms ] story [ 36.6 ms ] threadWe asked where this happened. He had done his work on an EC2 instance named `temp-worker-5`.
We asked where the files on that EC2 instance were; maybe we could cleverly get them back off the disk. They were in `/tmp`. We had it mounted on tmpfs (in memory), and we were running tmpreaper to clean out `/tmp` every Sunday.
When presented with the sheer absurdity of doing work on an instance called `temp`, out of a directory called `/tmp`, he backed down...
I fear that we wouldn't be so lenient (or rational) on my team had an AI agent done this - we'd have to enable the AI, after all.
Why did they pick that insane option? Why was it available to them?
One of my co-workers in security always seems to think that just saying "No." is the job. If someone is asking to do something stupid you say No, and then you talk to them to figure out what the goal is and the n you help them to get to that place.
Unfortunately, that conversation never happened.
Worse, people have had to explain to an (attempted) politician why bins cannot serve as shelters for homeless people.
I won't list the awful ways this had already wrong before this guy had this idea, but they are described at the end:https://politicsandinsights.org/2020/02/06/meet-the-multimil...
I also once knew someone who kept all of their most important documents in the recycle bin on their windows machine, as they felt they’d be safe from anyone snooping.
Anyway I’m off to take my toddler to the grenades that look like cartoon characters museum.
Not an intern, but when I was a new hire on a trading desk I dropped the dividend database that we used to price options. Cost the desk a boatload before we stopped offering quotes, after which the day was a total loss. That’s the sort of fuckup that lands in the CEO’s report.
I understood what I did ex post facto, it was an honest—if stupid—mistake, I expected to get fired and instead was given everyone’s shit work for six months during which I learned a tonne. One of my tasks was redesigning the system so a single careless keystroke couldn’t do that during the trading day.
So yeah, kinda your fault for not being more ready, sure, but also maybe it's worth bringing these sorts of failure-case stories to attention to help keep the sales pitch in perslective.
The AI vendors are promising the moon with millions of dollars of advertising while warning "But of course be careful!" in a 42% grey 8pt font disclaimer at the bottom of the UI if you scroll down enough after you buy in and start using it.
The duplicity and misrepresentation invites people to let down their guard.
I also let a guy that had one beer drive me 5 min on a rural road.
Both were pretty stupid and in the reconsideration, I wouldn't do it again.
But having it run company servers is just next level.
It's like believing that cola zero is healthy, because it's sugar free. No, it isn't, should probably be illegal to sell and people only buy it because of millions spent on marketing.
Just today I used it to get a summary of osteopat's credentials (Frank in Trondheim, guy's a treasure), because my brother asked who I'm talking about, when writing about feeling great after treatment on family chat.
It took 2 more prompts to get that summary, because the bot wouldn't stop ranting about how some osteopatic techniques are not yet medically sanctioned.
Thanks again for pointing that out. I feel childish for not seeing that issue. When I get more time I'll set it up in a VM to get more separation from my system.
Source?
Agents doesn't cause new problems, they reveal existing weaknesses.
Proper working backups. Proper IAM setup. With agents things can go wrong so much faster, the margins for allowed systemic/tech debt are shrinking.
And if you trust "AI" marketing, it will take all our jobs and do them better than us.
You're probably a regular tech news reader and are aware of the limitations, while they aren't.
A production company should 100% have regular backups, even moreso than the average developer.
https://www.aftonbladet.se/nojesbladet/a/31oE8A/the-challeng...
I know google has some manual process (forms) to request "undeletion" of personal Google Drive files.
It very probably was, but it just reminded me of this oversight failure mode.
[1] https://thedailywtf.com/articles/the_brillant_paula_bean
This problem predates agent based work. Anything an AI agent can accidentally do, malware could purposefully do too, as could a sufficiently careless human (one of the things my backups are intended to protect me from, is me!).