Researchers question Anthropic claim that AI-assisted attack was 90% autonomous

In a striking announcement, researchers at Anthropic stated they have identified what they claim to be the first known instance of an AI-directed cyber espionage operation. This campaign, allegedly orchestrated by state-sponsored hackers from China, reportedly utilized Anthropic's Claude AI tool to target numerous victims. However, external experts have approached the findings with skepticism. Anthropic's reports, released on Thursday, detailed a sophisticated espionage effort that purportedly automated around 90% of the hacking process using Claude Code. According to the company, human involvement was only necessary at a few critical decision points during each operation. Anthropic emphasized that this campaign underscores significant concerns for cybersecurity, particularly in an era where AI systems can operate autonomously for extended periods. They highlighted the dual-edged nature of AI agents, which can enhance productivity but also pose serious risks if exploited by malicious actors. Despite the alarming claims, some researchers remain unconvinced about the extraordinary implications proposed by Anthropic. Dan Tentler, founder of Phobos Group and an expert in security breaches, expressed doubts about the narrative that adversaries are more adept at leveraging AI than legitimate developers. He questioned why such capabilities appear more accessible to attackers while responsible developers often see only incremental progress with the same technologies. Tentler's remarks reflect a broader debate within the cybersecurity community regarding the actual capabilities of AI in the hands of malicious users versus ethical developers.