GPT-6 Astra finds 35,000 new vulnerabilities in SoftBank's own systems

What's going on

Junichi Miyakawa, CEO of the Japanese telecom company SoftBank, was unusually candid about the state of IT security when he presented results for the second quarter of fiscal 2026 on November 9. According to CNET Japan, he said that "to be honest, lately I've been spending almost all my time on countermeasures against cyberattacks." He also apologized for the ransomware attack on subsidiary IDC Frontier. Ransomware is extortion software that encrypts data and only releases it again in exchange for a ransom.

The headline number: when SoftBank ran OpenAI's top model GPT-6 Astra over its systems, the model found 35,000 previously undiscovered vulnerabilities. That came shortly after the company had already closed 10,500 flaws.

More attacks, and faster ones

Miyakawa says reconnaissance against SoftBank systems has risen sharply. Reconnaissance means probing systems for weak spots, much like burglars rattling doors at night to see which one is unlocked. Last year SoftBank counted around 300 million such probes a month on average. Recently that figure has roughly doubled. Since August, another 100 million have been added every month.

The speed is going up too. According to Miyakawa, the attack on IDC Frontier took just seven minutes. The traditional sequence of getting an alert, calling a meeting and making a decision is far too slow for that. So SoftBank has flipped the order: cut the connection first, then analyze and respond.

Miyakawa blames AI-assisted attacks. Since around May or June there have been AI models built specifically for cyberattacks, he said, and in July an open-weight model of this kind appeared, meaning its weights can be downloaded freely. Since then new models have been coming out almost every week. Every time SoftBank points a new model at its own systems, it turns up vulnerabilities nobody had found before.

From 10,500 to 35,000 flaws

SoftBank also uses AI for defense. Since July the company has offered a service called "Patching as a Service," built on OpenAI models, that helps businesses with everything from finding vulnerabilities to installing patches, the software fixes that close them. As of October 30, the service had 7,670 contracted customers.

SoftBank used the same process internally. On June 15, around 700 in-house systems were scanned. The result was 10,500 vulnerabilities, all fixed within about three months. Then came GPT-6 Astra, which OpenAI released as its flagship model in September. The second scan found 35,000 more flaws, over three times as many as the first round.

"To be honest, we are completely exhausted."

That's how Miyakawa described the situation. He expects the model coming out next week to find new vulnerabilities yet again, and said the work would go on. According to the report, he didn't say how severe the 35,000 findings are or how many of them can actually be exploited.

The lesson from the IDC Frontier attack

The scope has since been widened to more than 1,800 systems, including open-source software. That's software whose source code is public and which many companies use as building blocks. The reason is the IDC Frontier case. The in-house components there had already been patched, but the attackers got in through an open-source component that had never been scanned. Miyakawa said he had instructed teams from the start to examine open source thoroughly as well, "but the teams on the ground put it off." Further CNET Japan reports say the attack on the IDCF cloud hit 495 companies and municipalities, and data recovery is considered "difficult." The cause was reportedly a flaw in the authentication system.

SoftBank says the emergency review of its systems, including those of group subsidiaries, is complete and countermeasures have been brought forward.

The bottleneck: human review

The process has three steps that run in a continuous loop: find vulnerabilities, apply patches, then check that everything still works. The first two are largely automated. The third isn't. Humans have to verify that several interconnected systems still work together correctly after a patch, and that takes a lot of time. Miyakawa wants this step automated within the current fiscal year, but said he "can't give a firm date yet."

He compared the current situation, with human-built systems and AI side by side, to a transition period when human-driven and self-driving cars share the road: "Right now we just have to hang in there."

Markets ride the AI mood

Meanwhile, shares of SoftBank Group, the parent company with stakes in Arm and OpenAI among others, rose 11.22 percent on Monday, according to boerse-express.com. The site puts this down to a broader AI euphoria in Asia. It says Korea's Kospi gained 4.6 percent, driven by Samsung and SK Hynix in connection with GPT-6 Astra. The author also warns of annualized volatility of 77 percent and financing risks in the sector. The jump in the share price has little direct connection to Miyakawa's remarks.

Analysis

These numbers give you a rare, concrete look inside a major corporation. AI is finding security flaws faster than anyone can fix them. That cuts both ways. For defenders, it's good to find the holes before anyone else does. But attackers have access to similar capabilities, and according to Miyakawa they now even come as a freely available model. Whoever patches faster wins.

The numbers deserve some skepticism. They come from the company itself, and it's unclear how serious the 35,000 findings are, how many are false positives and how they were counted. SoftBank also sells exactly this kind of AI-powered patching service, so the dramatic figures double as an ad for its own product. Still, the IDC Frontier case shows the danger is real, and that vulnerabilities often sit where nobody is looking. The message for other companies is clear. If every new model generation means a new round of scanning, you need to plan staff and processes for it, especially for human review after patching.