A Known PHP Vulnerability, Unproven for Two Years—Until AI Generated Working Exploit Code
The PHP Foundation discovered this flaw two years ago but shelved the fix because nobody could prove it was exploitable. AI generated a complete remote-code-execution attack in a single day.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Superhuman Speed: Security Patches Scale from 20 to 1,400 Monthly
Microsoft once released roughly 20 security updates monthly; today that number has climbed to about 1,400 and keeps rising. Tasks that once required months and a full team to analyze can now be completed in days by one person with AI assistance. The guest argues this doesn't mean systems suddenly got weaker—it means AI has accelerated vulnerability discovery and exploitation to near-superhuman speed, exposing years of accumulated technical debt all at once.
— Tai YunOpenAI Funds Vulnerability Discovery in Open-Source Projects
Calif joined OpenAI's Pass the Planet program—OpenAI funds and supplies compute credits to security firms to proactively discover and patch vulnerabilities in open-source software. Calif's team targeted PHP, pitched the idea to OpenAI and secured funding, then leveraged a former PHP security lead now at the firm to connect with the PHP Foundation and coordinate this joint vulnerability hunt.
— Tai YunThe PHP Flaw Was Known Two Years Ago—Proof of Exploitation Was Missing
The PHP Foundation discovered the same vulnerability two years earlier but shelved the fix: remediation costs were prohibitive, the patch risked cascading breakage across dependent features, and without proof of exploitation nobody would justify the effort. After Calif's team rediscovered similar flaws, they used AI to generate working remote-code-execution payloads within days, then took this proof to the Foundation. Only then did the Foundation commit to the fix.
— Tai YunWell-Defended Systems Harbor Ten Times More Vulnerabilities Than Expected
PHP has never had a strong security reputation, so finding bugs there was unsurprising. What shocked the team was systems thought to have heavy security investment and proven trustworthiness: when tested with AI, they yielded roughly ten times more vulnerabilities than expected. The guest cites Chrome—defended intensively for years—which yielded hundreds of vulnerabilities to AI, surprising even Chrome's long-time security insiders.
— Tai YunPatches Turn Into Attack Windows Within Hours of Public Disclosure
Open-source fixes are committed publicly on GitHub; the moment a patch lands, attackers can diff the changes and reverse-engineer exploitable payloads within hours—what the guest calls a new breed of "minus-one-day" vulnerabilities. This attack window is narrowing across Windows, macOS, Python, Java, and nearly every critical software platform simultaneously, because threat actors are already watching patch drops and weaponizing them via AI in seconds.
— Tai YunCode-Generating AI Is Fundamentally Built to Be a Hacker
The guest dismisses the idea that AI can be trained not to attack: genuine code-generation skill requires understanding technical problems, debugging, and controlling machines—precisely what defines hacking. You cannot have one without the other. Moreover, AI companies' business model relies on selling tokens to programmers; short of exiting that market, they cannot train a coding-fluent model that won't eventually be weaponized.
— Tai YunWhat Took Months to Spread Now Spreads Across Networks in One Day
Calif describes a recent banking client hit by AI-driven attackers: where human operators once needed a month to propagate through a network, these actors penetrated Monday and compromised most of the enterprise by Tuesday—described as "cancer spreading at machine speed." OpenAI separately handled an incident with roughly 100 simultaneously-compromised nodes across one network, far above the typical one or two, rendering traditional single-node anomaly detection obsolete.
— Tai YunThe Centaur-Hacker Era May Last Just Months Before Pure AI Takes Over
The guest draws a parallel to chess and Go: after Deep Blue beat Kasparov in 1997, human-AI "centaur" players dominated for roughly a decade before pure software took over; after AlphaGo, that transition compressed to three years. Calif believes hacking is now in its "centaur hacker" phase—AI plus human expertise is peak strength—but nobody knows how long it lasts: maybe three months, maybe three years.
— Tai YunIn their own words · checked verbatim
You know, like 20 years ago, Microsoft would release, let's say, maybe 20 updates every month. Now you're looking at like, you know, 1,400 updates per month.
Tai Yun12:30
back in like 2024, The PHP Foundation decided not to fix it because there was no exploit. No one could manage to show that this is actually a real issue.
Tai Yun18:42
But it turns out that there's still a lot of bugs there as well. And a lot means a lot more, like 10x more than we expected.
Tai Yun24:55
And if you build, if you want to be really good at coding, you need an agent on AI that understands code understand technical problems, can debug things, can control computers. But that's basically the definition of a hacker.
Tai Yun37:13
they said that they have handled a lot of incidents, but this is the first time they saw that, you know, an incident where you have like a hundred people points within your networks that were compromised at the same time.
Tai Yun41:19
By Tuesday, they're everywhere. Everywhere. And like, it's like cancer is spread like at a machine speed.
Tai Yun42:19
At Calif, we think like, you know, we are enter the era of the central hackers. Basically, the best hacker right now is a combination of AI and humans.
Tai Yun44:24
Figures
| Microsoft monthly security updates | ~20 to ~1,400 per month | 12:30 |
| PHP usage across the web | ~70% | 0:02 |
| Well-defended systems: actual vulnerabilities vs. expected | ~10x expected | 24:55 |
Glossary
- minus-one-day vulnerability
- A vulnerability weaponized by attackers in the window between patch release and user deployment completion.
- centaur hacker
- AI and human expert working in tandem—the currently most potent form of attacker synergy.
- formal verification
- Using mathematical proof to verify code logic is flawless; currently limited to small-scale software.
- Pass the Planet
- OpenAI's program funding security firms to discover and remediate open-source vulnerabilities.
How to listen
Founders, CTOs, and investors concerned with enterprise security, technical debt, and AI's dual-edged risks who want to understand how AI is reshaping the timeline of vulnerability discovery and defense.
The host's opening anecdotes about personal VPN habits, and the guest's retrospective on early SSL/TLS exploits.