SoftBank CEO Miyakawa: AI-Driven Attacks Beyond Human Response

SoftBank had scanned its in-house applications and applied patches, but had not assessed the open-source component that was hit. The group is now extending checks to open-source software, going beyond the roughly 1,800 systems covered before.

Attacks are accelerating

According to the company, reconnaissance activity that precedes attacks has surged. It averaged about 300 million events a month in 2025 and has doubled to 600–700 million recently, rising by about 100 million each month since August. Miyakawa believes AI is being used in these attacks and linked the rise to open-weight models such as Moonshot AI's Kimi and Alibaba's Qwen, which began appearing around July. He said new models now arrive on a weekly cycle.

AI-based scanning keeps uncovering more flaws. In June SoftBank scanned about 700 of its systems with AI, found roughly 10,500 vulnerabilities and spent about three months patching them all. A rescan with OpenAI's GPT-6 Astra then found about 35,000. "Almost every time a new model comes out, we find new vulnerabilities," Miyakawa said, adding that he is exhausted but keeps patching.

Shift to automated defense

Because attacks are so fast, he said, "human hands may not be able to keep up." Instead of investigating and deciding a response in a meeting, SoftBank now cuts the connection first and handles the problem afterward.

Scanning and patching are already automated, but checking that systems run normally afterward still needs people. Miyakawa said he wants to automate that verification within the current fiscal period.