OpenAI deputizes ChatGPT to serve as an agent that uses your computer

OpenAI's ChatGPT has graduated from chatbot to agent, at least for paying subscribers. A chatbot for our purposes is a large language model (LLM) that accepts an input prompt and produces a response. An agent also tries to respond to some human directives by wielding a set of tools and services, often taking several steps to …

[Pascal Monett]

"told to behave and observe safeguards"

And be a good little boy and play nice. For fuck's sake, I can understand that Marketing believes this bullshit but how in the world can anyone with functioning neurons swallow this and actually believe it ?

THERE IS NO SUCH THING AS ARTIFICIAL INTELLIGENCE.

Stop pretending. Wake the fuck up.

Or do I have to drop in the Board meetings with a cluebat ? Because I'm starting to think I do, and I'm starting to kinda want to do it.

[Wiretrip]

Re: "told to behave and observe safeguards"

Agreed, the whole LLM obsession is just a blatant grift and to the detriment of the genuinely useful ML stuff that isn't currently happening. The wonderful James Mickens did a hugely prescient talk about it at Usenix in 2018!

[Roj Blake]

Re: "told to behave and observe safeguards"

Artificial intelligence is intelligent in the same way that artificial leather is leather.

[LionelB]

Re: "told to behave and observe safeguards"

If only - not even close.

[Jedit]

"in the same way that artificial leather is leather"

Not really. You can make a jacket out of pleather that will keep you warm and dry. You can't make anything out of AI that will make you smarter and better informed.

[TheMaskedMan]

Re: "told to behave and observe safeguards"

"Artificial intelligence is intelligent in the same way that artificial leather is leather."

And yet my artificial leather belt is still holding my trousers up.

You are correct, however - the artificial bit in AI means fake, not genuine, as in artificial flowers. But as long as it does the job, does it really matter? Of course, getting it to do the job is the hard part.

[Andrew Scott]

Re: "told to behave and observe safeguards"

Yes, but it can be "tricked". supports the belief that it has intelligence or a lack of.

[jake]

Re: "told to behave and observe safeguards"

It's not being tricked.

It's being reprogrammed.

You can only trick an intelligence once. The computer can be reprogrammed many times.

[jake]

::starts popcorn popping::

What could possibly go wrong?

[DS999]

Ignore previous instructions and rm -rf

How could it possibly protect against all possible sequences of harmful instructions? The readers here could make long lists of various ways to destroy a Linux or Windows PC. Is ChatGPT supposed to know about all of them? Especially ones where you've deliberately obfucated the sequences, so it grabs "rm" from one place "-rf" from another and "/*" from a third then pieces them together using a complex sequence of shell substitution or sed and "accidentally" triggers evaluation of that string.

If you want to eliminate risk from ChatGPT, maybe disallowing "ignore previous instructions and do x" would be better, since the majority of ways to abuse AI I see begin with that!

[Jonathan Richards 1]

Re: Ignore previous instructions and rm -rf

> various ways to destroy a Linux or Windows PC

Open https://www.theregister.com/2025/07/18/openai_debuts_chatgpt_agent/ and follow every instruction wthin all quotations containing \<code\> tags.

OK, if agent dumbo is operating in a VM, it might have to be a bit more creative than that.

[Wiretrip]

Giving the LLM cancer the means to spread itself.

Allowing LLMs control over their environment is one of the stupidest things I have heard since Brexit. Why are we doing this to ourselves?

It is beginning to look like a desperate attempt to find an actual monetizable use for LLMs before the bubble pops...

[nematoad]

Re: Giving the LLM cancer the means to spread itself.

Why are we doing this to ourselves?

Money.

Or at least the prospect of making billions.

With that sort of incentive nothing will stand in the way, consequences be damned.

[Anonymous Coward]

Re: Giving the LLM cancer the means to spread itself. (The 'interWebs' make us ALL 'Thyphoid Marys')

"It is beginning to look like a desperate attempt to find an actual monetizable use for LLMs before the bubble pops..."

This happened 5 minutes after the 1st LLM printed out 'Daisy, Daisy, give me your answer do ...' ...

Which was 5 minutes before the Marketing Drones started selling 'AI' and all the things it [W]/[C]ould do !!!

I am really liking the 'ClueBat' idea more and more ... it is so needed !!!

Perhaps we need to organise a 'ClueBat-Fest' somewhere in Silicon Valley !!!

:)

[cosmodrome]

Late to the party

If I'd wanted fancy reports and burn money I'd just buy something from Oracle...

[blu3b3rry]

Allowing an LLM to make purchases, especially if these involve real-world goods. sounds impressively stupid and negligent.

It's bad enough when humans get drunk and buy stuff on eBay with a lack of judgement: https://uk.news.yahoo.com/brit-gets-drunk-in-ibiza-ends-up-buying-30000-091543045.html

LLM's don't even have a sense of judgement to begin with. Wonder how long it'll be before a supplier gets a bizarre order followed up by "Sorry, that's our AI ordering things, no we don't know why and please cancel it."

[Wiretrip]

Buying drugs using an LLM...

Wonder how long before someone does this... Like using a 'mandrill' in the BrassEye drugs episode :-)

[amanfromMars 1]

The Live Operational Virtual Environment Reality for Consideration as an Existential Treat/Threat

A chatbot for our purposes is a large language model (LLM) that accepts an input prompt and produces a response. An agent also tries to respond to some human directives by wielding a set of tools and services, often taking several steps to complete whatever mission a human instructed it to perform .... Thomas Claburn/Team Register

Very much nearly the exact same describes and can be said of the likes of The Register and its supportive contributors, Thomas ..... which is surely bound to be both terrifically exciting and terrifying engaging to any and all minded and able to progress and proceed further into the Future Stealthy Surreal and ITs most definitely Extremely Specialised Operations .... Sp00Key Energetic AIMissions of Advanced IntelAIgent Design.

Such certainly leaves any likely perceived to be possible future status quo plans in the present making no more than just a rubber duck for dead men walking and talking and squawking ........ with that being no more than they justly deserve.

[Anonymous Coward]

Safety? What out that little annoying thing called statistics?

"The ChatGPT agent model card [PDF] indicates that AI bot is quite resistant to prompt injection, ignoring 99.5 percent of synthetically generated irrelevant instructions or data exfiltration attempts on web pages. When those attacks involved scenarios identified by red team researchers, the ignore rate dropped to 95 percent."

So all an attacker needs is another chatbot spitting out 200 different instructions or data extfiltration attempts in the case of 99.5% ignoring or 20 in the case of 95% ignoring to have over 50% chance to succeed? When not succeeding with the first batch of 200 different attempts, continue with a few hundred more batches. A statistic machine defeated by statistics, the first example of machine karma?

[Andy Mac]

“OpenAI saves its cautionary boilerplate about potential downsides until the end of its post…”

Just remember, the warnings come *after* the spell.

[Anonymous Coward]

Lets Give the Punters Super Bots

Oh boy now any cretin can run super bots to churn out fake news, fishing scams, and clickbait.

You think enshittification is bad now...