PHP2Shell: two-days-to-find, two-months-to-fix PHP RCE vulnerabilities aided by AI
Core Highlights
Two long-dormant PHP remote code execution vulnerabilities, collectively called PHP2Shell and tracked as CVE-2026-109618 and CVE-2026-109619, were disclosed with CVSS scores near the maximum. Most notably, the bugs were found by an AI model, GPT-5.6 Sol, while assisting security research, and OpenAI subsequently released a dedicated detection skill. A tug-of-war of two days to discover, two months to fix puts the complexity of modern vulnerability disclosure on full display, and it is a story with implications well beyond PHP itself.
What Happened
The whole process played out like an absurd drama. The vulnerabilities leaked early, a demo video circulated on the open web, and multiple government websites preemptively went offline to avoid being compromised. It was only at 10 p.m. on the 21st that PHP's official patch went live. The author reconstructed the full timeline from leak, video spread, government emergency takedown, to final fix, and pointed out that the source of the storm was AI-assisted discovery, a detail that reshaped how the episode is understood by the security community.
Technical Details
Both vulnerabilities are remote code execution, meaning attackers can run arbitrary code on the server through flaws at the PHP engine level. Since PHP powers roughly 70% of the world's websites, the impact is extremely broad and touches everything from small blogs to large government portals. The fix involved substantial changes to the PHP engine and required repeated coordination among maintainers, researchers, and security vendors, which is why the fix took two months. OpenAI's detection skill helps sites check whether they are affected, turning a scary unknown into a checkable condition that defenders can act on immediately.
Compared With Alternatives
Traditionally, vulnerabilities were found through manual audits or fuzzing, long cycles and narrow coverage. This time an AI model pinpointed a severe flaw first during assisted analysis, showing that large models are becoming a force multiplier for security research. Yet AI-found vulnerabilities still must go through coordinated disclosure, no different from human discoveries, and the two-day discovery, two-month fix asymmetry is the real lesson.
Industry Impact and Use Cases
For operations and developers, the most practical action is to upgrade the PHP version quickly and deploy official or vendor mitigations. For the industry, AI-assisted vulnerability discovery will become routine, but the accompanying discover-coordinate-disclose mechanism must keep pace, or a single leak can escalate into a global incident before a patch is even ready.
Practical Notes
If you run PHP, treat this as a patch-now item. Verify your version against the fixed release, and if you cannot upgrade immediately, apply the vendor's virtual-patch or Web Application Firewall rule as a stopgap. Use OpenAI's detection skill or an equivalent scanner to confirm whether your deployment is exposed, and re-scan after any change to your stack.
What to Watch
Expect more AI-assisted disclosures like this. The two-day discovery, two-month fix asymmetry will keep biting because discovery is accelerating while coordinated patching is not. Monitor your dependency and engine versions the way you monitor uptime, and build an emergency-patch runbook before the next one lands.
Takeaway
PHP2Shell is a turning point that shows AI can now find world-scale flaws faster than the ecosystem can fix them, and it raises the stakes for how responsibly discovery and disclosure are handled across the entire software supply chain.
Background
The disclosure is a milestone because it is among the first world-scale vulnerabilities credibly attributed to AI-assisted discovery, and it changes the threat model in a way defenders have not yet adapted to. From now on, the race is not only against human researchers but against models that can scan an engine for flaws in days, and that compresses the window between discovery and exploitation.
Where to Look Next
Follow whether CVE-2026-109618 and 109619 prompt a broader audit of other ubiquitous runtimes, because if similar AI-assisted finds surface in other languages or frameworks, the two-day-discovery, two-month-fix gap becomes a systemic risk the industry must engineer around with emergency patching baked into the release process.
Closing Thoughts
PHP2Shell should be read as a warning that the discovery side of security is being automated faster than the response side, and closing that gap is now a prerequisite for running any widely deployed runtime safely.
Final Note
The uncomfortable conclusion is that AI is now on both sides of the vulnerability equation, discovering flaws faster than humans ever could while the response machinery remains stubbornly human-paced. Until coordinated disclosure and emergency patching are funded and automated at the same scale as AI-assisted discovery, every widely deployed runtime carries a latent risk measured not in days but in the gap between a model's insight and a maintainer's release. That gap is the new attack surface, and it is widening.
That gap is the new attack surface, and closing it is now everyone's job.
The Coordination Gap
What the timeline really exposes is a mismatch of clocks. Machine-assisted discovery now runs at the speed of inference, while coordinated disclosure, maintainer review, and emergency patching still run at the speed of human organizations. That gap is where risk accumulates: the window between a flaw becoming knowable and a fix becoming installable. Closing it requires treating disclosure infrastructure as seriously as discovery research, including funded maintainer time, pre-agreed emergency channels, and automated detection that defenders can run as easily as attackers can scan. For operators, the immediate action is mundane but urgent: inventory exposed PHP services, apply the vendor patch, and verify with the published detection tools rather than assuming a default installation is safe. In a world of AI-assisted finding, the defensive side has to automate its response or accept a permanently widening exposure window.