Researchers exploit OpenAI's Atlas by disguising prompts as URLs

Researchers have found more attack vectors for OpenAI's new Atlas web browser – this time by disguising a potentially malicious prompt as an apparently harmless URL. NeuralTrust found that Atlas's "omnibox" (where URLs or search terms are entered) has potential vulnerabilities. "We've identified a prompt injection technique …

[Irongut]

Trusted user input

Well there's your basic programming failure right there. As Mrs Tables will tell you, you never trust user input or her son Bobby.

These "AI" companies are making basic programming mistakes that we've known about for decades.

[elsergiovolador]

Re: Trusted user input

But people who know these basics have retired or got fired.

[Alex Stuart]

Not that long ago

..I remember a common response to concerns about AI going rogue was along the lines of "well of course we'd only run it on an airgapped system, we obviously wouldn't let an intelligence we didn't fully understand just have carte blanche access to the Internet, that'd be insane lol, we've all read the sci-fi stories, come on"

That aspiration does not seem to have survived the realisation that $loads of ad and subscription money is on the table.