DARPA Broadcasts AI Cyber Problem Finalists

Seven groups superior to the finals after the DARPA AI Cyber Problem (AIxCC) semifinal competitors at DEF CON 32 final week. The groups had been every awarded a $2 million prize and can take part within the remaining competitors in August 2025.

A partnership between DARPA and the Superior Analysis Tasks Company for Well being (ARPA-H), the AIxCC competitors had groups design synthetic intelligence (AI) techniques to safe open supply infrastructure software program that’s used throughout the non-public and public sectors in areas like finance, utilities, and well being care. These techniques are weak to cyberattacks as a result of they’ve massive assault surfaces and lack safety instruments at scale. Certainly, many have been latest targets, highlighting the pressing want to guard vital infrastructure.

For the competitors, groups had been requested to develop cyber reasoning techniques that might discover and repair vulnerabilities in a set of “problem initiatives” that had been designed by AIxCC consultants. Almost 40 groups submitted techniques. The rivals’ techniques found 22 distinctive artificial vulnerabilities and patched 15 of them. The techniques discovered 11 distinctive patches for C-based challenges and 4 for Java-based challenges. One real-world bug was present in SQLite3 and has been disclosed.

“In true DARPA style, we did not know if our speculation could be confirmed after we launched this program. Now we have seen that AI techniques are able to not solely figuring out but in addition patching vulnerabilities to safeguard the code that underpins vital infrastructure,” stated Andrew Carney, program supervisor for AIxCC, in an announcement.

The next groups will advance to the ultimate:

Groups have one 12 months to develop their techniques earlier than subsequent 12 months’s remaining competitors. AIxCC will distribute $29.5 million in prize cash to the groups judged to create the best techniques, and winners should launch their techniques as open supply software program after the competitors.