AI Security Needs a Major Overhaul: Experts

Cybersecurity experts at DEF CON, the world’s largest hacker conference, have raised concerns that current AI security methods are insufficient and require a complete rethink. According to the conference’s first-ever “Hackers’ Almanack” report, created in collaboration with the University of Chicago’s Cyber Policy Initiative, existing security approaches fail to effectively address emerging threats.

The report critiques “red teaming,” where security specialists attempt to find vulnerabilities in AI models. Sven Cattell, head of DEF CON’s AI Village, argued that public red teaming is ineffective due to fragmented AI documentation and inadequate evaluation methods.

During DEF CON, nearly 500 participants tested AI models, with even newcomers successfully discovering vulnerabilities. Researchers now recommend adopting a standardized framework similar to the Common Vulnerabilities and Exposures (CVE) system used in traditional cybersecurity since 1999. This would provide a structured way to document and fix AI weaknesses instead of relying on occasional security audits.