OpenAI Nears Launch of AI Agent Tool To Automate Tasks For Users

An anonymous reader quotes a report from Bloomberg: OpenAI is preparing to launch a new artificial intelligence agent codenamed "Operator" that can use a computer to take actions on a person's behalf (Warning: source may be paywalled; alternative source), such as writing code or booking travel [...]. In a staff meeting on Wednesday, OpenAI's leadership announced plans to release the tool in January as a research preview and through the company's application programming interface for developers [...]. The one nearest completion will be a general-purpose tool that executes tasks in a web browser, one of the people said.

OpenAI Chief Executive Officer Sam Altman hinted at the shift to agents in response to a question last month during an Ask Me Anything session on Reddit. "We will have better and better models," Altman wrote. "But I think the thing that will feel like the next giant breakthrough will be agents." The move to release an agentic AI tool also comes as OpenAI and its competitors have seen diminishing returns from their costly efforts to develop more advanced AI models.

Great attack vector!

[gweihir]

And if the "AI" hallucinates, it may just delete all your data without even any need for an attacker.

Does anybody here think this is a good idea?

Re:

[starworks5]

Its not a a good idea for users and it is a huge security vulnerability, but they have to look like they are not lagging behind Anthropic, who released a similar tool. That being said, its not a bad thing that an AI agent can interact with a desktop, i.e. an ai agent that can use MS Paint in a sandbox, the problem is that people are going to deploy this tool in a insecure manner.

Re:

[gweihir]

the problem is that people are going to deploy this tool in a insecure manner.

People are always doing that, because people are not experts. Hence there are whole classes of tools that regular people either cannot get or it is simply not marketed to them.

Re:

[zlives]

can they create an agent to help fix issues with browser and outlook plugins, i am assuming this will be a plug in

Re:

[gweihir]

Unlikely. There will not be enough training sdata. This crap cannot solve problems. It only can do some statistical adaptation and mix of solutions it has seen before in its training data.

Re:Great attack vector!

[nightflameauto]

And if the "AI" hallucinates, it may just delete all your data without even any need for an attacker.

Does anybody here think this is a good idea?

While it is an attack vector, what the management types are seeing is a glorious way to train a person's replacement. If the AI agent doesn't cause any clusterfucks? Time to take that task off the human's plate. Nobody gives a flying fuck about security issues until they actually cause damage. Then it's a, "We sincerely apologize" moment for some middle manager who'll get fired, while the upper management continues to scramble to find new ways to break security through automation.

Re:

[gweihir]

Nobody gives a flying fuck about security issues until they actually cause damage.

Damage from security problems for example in Germany in 2023: 2600 EUR per person (!). This is very likely a significant underestimation. Insecure software has become a major economic factor. Unreliable software adds damage on top of that.

Re:

[nightflameauto]

Nobody gives a flying fuck about security issues until they actually cause damage.

Damage from security problems for example in Germany in 2023: 2600 EUR per person (!). This is very likely a significant underestimation. Insecure software has become a major economic factor. Unreliable software adds damage on top of that.

I don't disagree at all. How do you convince decision makers of that without "proof" when they think they can save pennies by throwing security out the window?

Re:

[gweihir]

Look at the EU KRITIS initiative. The damage has gotten too high to leave this to "decision makers" so these incompetent and greedy fucks are now getting regulated. I guess such an idea will not fly in the US.

Re:

[BringsApples]

Security? What's that? Since all data is in the cloud, aka in datacenters, then all of the data is centrally located. Since all of the data is centralized, that means that all intelligence is centralized. Now all we need is a Central Intelligence Agency and they can make sure that no one uses said intelligence for the bad. ...oh, wait.

Re:

[Mirnotoriety]

> And if the "AI" hallucinates, it may just delete all your data without even any need for an attacker.

Or sign you up to Dignitas.

Re:

[AleRunner]

Well, until you said that signing people up to dignitas is often the perfect solution for security problems, that wasn't in the training set so it wouldn't have done it. Now you've blown it.

Re:

[gweihir]

And that does what? Dignitas is not an assasination service...

Re:

[arglebargle_xiv]

It's OK, it's been nearly a week since AI slop peddler Altman got to use Slashdot as his personal advertising platform, this is just filling the quota.

booking travel

[zlives]

i mean if their AI can book travel based on my requirements
cost, time of travel, time to travel, stop overs, stop over location preference, air lines preference,

i could use it. does it come with a big FU warning that this automation is not automation and the end user is responsible for all mistakes? ie autopilot

Re:

[Fly Swatter]

For business it would be far more efficient to just send the AI 'over' to wherever and do whatever and avoid air travel all together. Oh wait, that's your job? Oops.

Re:

[narcc]

You have some seriously misguided ideas about what this technology is capable of doing. Did you honestly believe Eric Yuan when he said 'it's just down the stack'?

Re:

[Fly Swatter]

A) sarcasm is lost on you, B) this tech is just a big database lookup, any tasks that need to be reliable will be mostly scripts; chat bot scripts are new again! C) I was criticizing the wastefulness of air travel business trips as a job.

Al doesn't exist yet.

Re:

[narcc]

this tech is just a big database lookup

No, it's not. Where did you get that idea?

I was criticizing the wastefulness of air travel business trips as a job.

Context matters.

Re:

[gweihir]

does it come with a big FU warning that [...] the end user is responsible for all mistakes?

Obviously. If it books you first class to fly DC to Seattle via Hong Kong and Sydney, that is your problem.

I'm sorry Dave

[Tony Isaac]

I'm afraid I can't do that.

Not interested in having it book travel, but

[Tony Isaac]

It would be nice if it could do things like adjust obscure control panel settings, like "Change the power settings so that my computer never goes to sleep when plugged in." That's what I *thought* Microsoft's Copilot might do, but no, it just gives you step-by-step instructions that are _sometimes_ right.

Before anyone says "privacy"

[Iamthecheese]

Not enough people are aware how easy it is to run an LLM locally, and the fact that a midrange GPU gives answers nearly as good as the higher end services (if not so close to the very top tier) and better than most of the free services. This definitely can be run locally, and there are a lot of people working on creating open source locally runnable LLM everything.

Great!!

[gtall]

When will it take out my garbage and help me use the urinal?

OpenAI - Find Job for us so we can eat!

[BrendaEM]

Look forward to that depression-threshold 10% unemployment.