An AI model found an unknown Google bug in 13 minutes for $4.24
Xint tells that the cost to find one unknown vulnerability went from $350 and no result to $4.24.
Xint, a security firm, tested seven new AI models on one unknown vulnerability in the V8 engine of Google. No model saw it in training, and in February the best models found nothing after 25 hours and $350. In September, one model found the vulnerability in 13 minutes for $4.24 with its first guided try.
The test
Each model had to examine a small set of source files that had the vulnerability. The vulnerability lets an attacker start code from a remote location. Two models came from OpenAI and three from Anthropic. The other two were open source.
The numbers
Only three of the seven models found the vulnerability. All three came out in July or after. Claude Opus 5 found it on the first try, but it cost $43.40. The two open source models did not find it.
What is not known
Xint tells that expertise is important, because some models had more guided tries. The test used one vulnerability in a small set of files. Xint tells that amateurs with unguided models cannot hack systems at this time.
Insight: Cheap discovery meets cheap exploits
Xint puts one find at $4.24. Forescout tells that its AI tool found 99 vulnerabilities in 90 hours for $1,200 to $2,300. Anyday Security opened a free catalog of more than 200 exploits that AI made. Discovery and use are cheap, and your patch date sets your risk more than attacker skill.
Sources
Posted