It Was Always Going To Be This Way, We Tried To Warn The World: Anthropic's Deputy CISO Jason Clinton On Recent Cyberattacks

Internet security teams are dealing with one of their hardest weeks in a long time. Jason D. Clinton, Anthropic’s Deputy Chief Information Security Officer, has a warning: these problems were always going to happen, and things will get worse before they get better.

Clinton shared his warning on X while security teams around the world were still working to install emergency fixes for two serious security problems in PHP. At the same time, Indian company eMudhra was dealing with the results of a compromised digital certificate key. The incident showed how one mistake involving a security key can create fears about cybersecurity around the world.

Clinton wrote that although the past four days had been very bad, he expected things to get much worse before improving. He told security leaders to use the crisis as a chance to ask their company boards for help immediately. He said they should ask for support with basic security measures, not AI tools. He also said organizations needed to get through about 18 months of these problems.

Another X user shared Clinton’s post and blamed Anthropic and similar companies for helping create the problem. The user accused them of creating a monster and then asking everyone else to help clean up the mess. Clinton answered directly, agreeing with the criticism. He said this situation had always been expected and that they had tried to warn the world.

A Growing Wave of Security Problems

Clinton has previously warned about what he calls a coming "vulnerability tsunami." He means a huge wave of software security problems caused by increasingly powerful AI models. These advanced models could become much better at finding, understanding, and using weaknesses in software to carry out attacks. They could do this faster than security teams can fix the problems. Clinton believes organizations should start using ways to separate important computer systems from dangerous connections right now. These include air-gaps, which keep systems physically disconnected from networks, and data diodes, which allow information to travel in only one direction. He argues that companies should not simply wait for software fixes that might arrive too late.

The events of this week have shown what this danger could look like. Cloudflare’s CEO said attempts to attack the newly revealed PHP weaknesses rose from thousands to millions within hours of the problems becoming public. Governments also acted quickly. Japan’s Prime Minister and the US Federal Bureau of Investigation (FBI) both publicly warned organizations to install fixes immediately. In a separate incident, cloud platform Railway reported a security breach that made it stop new software deployments because private customer information might have been exposed. Node.js also released fixes for a serious weakness in its Brotli decoder. This weakness could have allowed attackers to run their own computer instructions from a distance. The problem also led Vercel to quickly introduce emergency firewall rules to block dangerous activity.

Meanwhile, eMudhra’s problems have shown how much a security mistake can cost a business. After the company’s digital certificate key was compromised, it lost about ₹1,450 crore in stock market value in only two days. Its shares had already fallen by 20%, reaching the market’s lower price limit, after news of the breach became public. Security researchers studying what happened have suggested that the cause might have been a surprisingly simple mistake. They think there may have been confusion between 256-bit cryptographic keys and ordinary numbers from 0 to 256. This is still a theory, not a confirmed explanation. It would be the kind of basic security mistake that Clinton says companies must fix before AI-based security tools can help them.

This same week has also shown that powerful AI models can be used for both good and harmful purposes. OpenAI said its GPT-5.6 Sol model helped identify one of the serious PHP security weaknesses that caused so much concern. This shows how AI can help security teams check computer code for problems. But the same abilities could also help attackers find those weaknesses first and use them to break into systems.

The Next 18 Months: Working Toward 'Vulnzero'

People on X asked Clinton what he expected to change after the 18 months he mentioned. He answered by describing an idea he calls "vulnzero." He believes that as AI technology improves, the powerful reasoning abilities of today’s most advanced models, which he calls "cyber mythos-class" abilities, will eventually become available in much smaller and cheaper models. This is based on scaling laws, which describe how AI abilities change as models and the technology behind them improve. Clinton expects these smaller models to become cheap enough that software developers everywhere can use them to check every line of code for security problems as they work. Each check could cost only a fraction of a cent. The goal would be to stop whole types of software bugs before the software is ever released.

This idea depends on the same kind of advanced thinking and problem-solving that Anthropic has shown in other research. For example, Claude recently worked on solving a 25-year-old mystery involving NIST cryptography. The mystery concerned where the starting values, called seeds, for ECDSA mathematical curves originally came from. Clinton believes this kind of careful, advanced technical reasoning will eventually become a normal part of the tools that all software developers use.

But until that happens, Clinton is giving security leaders much simpler advice. He says Chief Information Security Officers, or CISOs, should not rush to add unfinished AI security products while they are already dealing with serious attacks. Instead, they should use the attention that company boards are now giving cybersecurity to ask for money to improve basic protections. These include keeping software updated with security fixes, controlling who can access important systems, and dividing networks into separate sections so attacks cannot spread as easily. These basic protections may not sound exciting, but Clinton believes even the most powerful AI models cannot replace them during the coming months.

For Clinton, the current security problems are not surprising. Instead, they support what he has been warning about. As he told people who criticized him, this was always going to happen.