the internet could go DARK...

Summary

Wes Roth argues that AI is rapidly changing cybersecurity, making both attacks and defenses more capable. He points to reports of a Quest Diagnostics data breach, a security researcher’s account of thousands of automated attacks on a decoy server, and Microsoft’s report on a China-linked campaign in which attackers used AI models and reportedly reached cloud access with administrative privileges in about three hours. Microsoft said most attempts in that campaign were unsuccessful, but Roth warns that observed attacks may represent only a fraction of the activity. He also says some attacks on newly disclosed PHP vulnerabilities began before the public announcement.

Roth’s broader concern is that AI systems can identify and exploit software weaknesses faster and more persistently than conventional attackers. At the same time, AI agents are being used to carry out tasks such as shopping or financial activity, creating risks when agents are given access to accounts or personal data. He cites examples of vulnerabilities involving agents and data exposure, while noting that at least one issue he discusses appears to have been fixed.

Discussing Sam Altman’s comments that it may be impossible to patch every vulnerability, Roth argues that organizations cannot assume today’s AI capabilities are the limit. Models may soon uncover weaknesses that current defenses cannot anticipate. He interprets Altman’s remarks as pointing toward a different cybersecurity model—possibly one that uses AI continuously and proactively to defend systems.

Roth is skeptical that simply pausing AI development would solve the problem. Existing open-source models would remain available and could be modified, while criminals or rival states might continue developing capabilities even if companies in some countries stopped. He also worries that a pause could leave defensive efforts behind offensive ones. He cautions against turning the issue into a simplistic political fight over “good” and “bad” sides.

A more hopeful possibility, Roth says, is that AI tools become cheap and capable enough to test software extensively before release. He discusses Anthropic cybersecurity figure Jason Clinton’s estimate of roughly 18 months toward a “zero vulnerability” future, in which models help developers find and fix flaws at scale. Roth presents this as a possible long-term outcome, not a guarantee; he expects a turbulent period first, with cyberattacks potentially increasing.

The video’s headline warning—that parts of the internet could be shut down temporarily—is presented as a possibility, not a prediction that a shutdown is certain. Roth recommends basic precautions:

• Use unique, strong passwords.
• Enable two-factor authentication through an authenticator app or hardware key rather than SMS where possible.
• Consider keeping some cash available in case services are disrupted.

He emphasizes not panicking and frames these steps as sensible even if the feared disruption never occurs.

Presenter and Contributors

• Presenter: Wes Roth
• People cited or discussed: Sam Altman, Jason Clinton, Bernie Sanders, Andriy Karpathy, and Steven Alyssa
• Organizations and sources referenced: Quest Diagnostics, Microsoft, and Anthropic