Key Takeaways

  • OpenAI and Wiz found PHP2Shell, a remote code execution bug in PHP that exposed roughly 70 percent of the internet.
  • Major tech platforms and the PHP Foundation coordinated patches in secret for two months before a public leak triggered millions of automated attack attempts per minute.
  • An Indian root key provider configured their server with a key value of literally 4 instead of 328 bits, exposing baseline trust mechanisms.
  • Prince projects that within five years, automated agent web traffic will run one thousand times higher than human traffic.

The Secret Patch Window That Saved the Web

When OpenAI combined their models with independent security research, they found a remote code execution flaw sitting inside PHP. The bug, dubbed PHP2Shell, created an open door into roughly 70 percent of all web servers. Rather than sounding a public alarm immediately, OpenAI went straight to Cloudflare and the PHP Foundation.

“OpenAI originally reached out to us to let us know that they, in combination using their models but also some work that some security researchers had done, had found a vulnerability in a piece of software called PHP,” Matthew Prince said. That outreach kicked off an intense, two-month private coordination effort among internet infrastructure giants to patch core systems before attackers noticed.

Coordination bought time, but the clock ran out. Once an external leak surfaced, bad actors automated the exploit almost instantly. As Prince explained, “At this point, we're now seeing millions of attacks per minute across our network where people are trying to compromise servers.” The incident proves that zero-day defense has collapsed into a race between automated patch delivery and automated exploit discovery.

Human Stupidity Versus Automated Discovery

AI takes the blame whenever software breaks in public, but human shortcuts still trigger the most embarrassing security collapses. While analyzing certificate vulnerabilities, Prince pointed out an incident where a certificate authority in India failed basic cryptographic standards: “an Indian company that was running some of these root key servers, they instead of using 328 bits, they used literally the number four.”

Attackers did not need a supercomputer to bypass that certificate. They just needed someone to test the door. Prince rejects the idea that artificial intelligence makes systems inherently more dangerous. Software already sits full of structural flaws. Automated systems simply locate what humans missed.

“AI is not the bad guy here. AI is the good guy. The AI didn't create these vulnerabilities. AI found them,” Prince argued. “The question is like if you're living in a house that's got deadly monsters living in every closet, but you don't know they're there. Are you in a better place or a worse place than if you start going through the closets and finding the monsters and killing them?”

Defenders who use model-driven audits clean out closets before attackers buy access. Those who refuse to audit their own stacks leave the doors unlocked for automated scrapers.

The Shift to Machine Traffic

Every web economic model assumes humans sit behind the browser. That assumption will not survive the decade. Prince offered a simple projection: “if current trends continue in 5 years, automated traffic agent traffic is going to be a thousand times human traffic.”

If machine requests outnumber human clicks by three orders of magnitude, standard pay-per-click models, basic firewalls, and static rate limiters break completely. Web servers must distinguish between malicious exploit probes, legitimate consumer agents, and human sessions without throttling genuine revenue. Builders who do not retool their network edge for machine-scale interactions will spend their days triaging server downtime.

What to Do With This

Run an automated static analysis pass across your legacy dependencies this week using frontier LLMs. Do not assume your open-source libraries or third-party web dependencies are secure just because nobody has attacked them yet. Set strict, edge-level rate limits on any server endpoint running interpreter languages like PHP or Python before an automated scanner finds it for you.