Google’s AI-powered bug hunter has simply reported its first batch of safety vulnerabilities.
Heather Adkins, Google’s vp of safety, announced Monday that its LLM-based vulnerability researcher Large Sleep discovered and reported 20 flaws in numerous well-liked open supply software program.
Adkins mentioned that Large Sleep, which is developed by the corporate’s AI division DeepMind in addition to its elite group of hackers Undertaking Zero, reported its first-ever vulnerabilities, largely in open supply software program corresponding to audio and video library FFmpeg and image-editing suite ImageMagick.
On condition that the vulnerabilities will not be mounted but, we don’t have particulars of their influence or severity, as Google does not yet want to provide details, which is a normal coverage when ready for bugs to be mounted. However the easy proven fact that Large Sleep discovered these vulnerabilities is critical, because it exhibits these instruments are beginning to get actual outcomes, even when there was a human concerned on this case.
“To make sure top quality and actionable reviews, now we have a human professional within the loop earlier than reporting, however every vulnerability was discovered and reproduced by the AI agent with out human intervention,” Google’s spokesperson Kimberly Samra instructed TechCrunch.
Royal Hansen, Google’s vp of engineering, wrote on X that the findings show “a brand new frontier in automated vulnerability discovery.”
LLM-powered instruments that may search for and discover vulnerabilities are already a actuality. Apart from Large Sleep, there’s RunSybil and XBOW, amongst others.
Techcrunch occasion
San Francisco
|
October 27-29, 2025
XBOW has garnered headlines after it reached the top of one of many U.S. leaderboards at bug bounty platform HackerOne. It’s necessary to notice that typically, these reviews have a human during the method to confirm that the AI-powered bug hunter discovered a reliable vulnerability, as is the case with Large Sleep.
Vlad Ionescu, co-founder and chief know-how officer at RunSybil, a startup that develops AI-powered bug hunters, instructed TechCrunch that Large Sleep is a “legit” mission, on condition that it has “good design, folks behind it know what they’re doing, Undertaking Zero has the bug discovering expertise and DeepMind has the firepower and tokens to throw at it.”
There’s clearly quite a lot of promise with these instruments, but additionally important downsides. A number of individuals who keep totally different software program initiatives have complained of bug reviews which can be really hallucinations, with some calling them the bug bounty equal of AI slop.
“That’s the issue individuals are working into, is we’re getting quite a lot of stuff that appears like gold, but it surely’s really simply crap,” Ionescu beforehand instructed TechCrunch.
