Google’s AI bounty program pays bug hunters up to $30K

On Monday, Google launched a caller reward program dedicated specifically to uncovering bugs successful AI products. Google’s database of qualifying bugs includes examples of the benignant of rogue actions it’s looking for, similar indirectly injecting an AI punctual that causes Google Home to unlock a door, oregon a information exfiltration punctual injection that summarizes each of someone’s email and sends the summary to the attacker’s ain account.

The caller programme clarifies what constitutes an AI bug, breaking them down arsenic issues that usage a ample connection exemplary oregon a generative AI strategy to origin harm oregon instrumentality vantage of a information loophole, with rogue actions astatine the apical of the list. This includes modifying someone’s relationship oregon information to impede their information oregon bash thing unwanted, similar one flaw exposed previously that could unfastened astute shutters and crook disconnected the lights utilizing a poisoned Google Calendar event. 

Bug hunters person already raked successful implicit $430,000 during the 2 years since the company officially started inviting AI researchers to basal retired imaginable avenues to maltreatment AI features successful its products. 

Simply getting Gemini to hallucinate volition not chopped it. The institution says that issues related to content produced by AI products — specified arsenic generating hatred code oregon copyright-infringing contented — should beryllium reported to the feedback transmission wrong the merchandise itself. According to Google, that mode its AI information teams tin “diagnose the model’s behaviour and instrumentality the indispensable long-term, model-wide information training.”

Along with the caller AI reward program, Google besides announced connected Monday an AI cause that patches susceptible codification called CodeMender. The institution says it has utilized to spot “72 information fixes to unfastened root projects” aft vetting by a quality researcher.  

The $20,000 prize is awarded for rooting retired rogue actions connected Google’s “flagship” products Search, Gemini Apps, and halfway Workspace applications similar Gmail and Drive. Multipliers for study prime and a novelty bonus are besides available, which could bring the full magnitude up to $30,000. The terms drops for bugs recovered connected Google’s different products, similar Jules oregon NotebookLM, and for lower-tier abuses, specified arsenic stealing concealed exemplary parameters.