Millions of exploitation attempts targeted newly disclosed PHP vulnerabilities within hours of publication, according to Cloudflare analysis, as attackers used broadly available AI tools to move faster. The flaws, grouped under the name PHP2Shell, forced security teams and web administrators into urgent patching, yet the PHP Foundation reports no successful exploitation so far. The episode matters for business because AI shortened the path from disclosure to attempted attack, raising the cost of slow patching.
How researchers turned old PHP flaws into exploits
Researchers at security company Calif first identified PHP2Shell and built working exploits in early August, combining frontier AI tools with deep human knowledge of PHP behavior and weaknesses. Part of the work ran inside OpenAI Patch the Planet program. Some underlying issues had been flagged as early as 2024 but were then considered too complex to exploit. After Calif, researchers at OpenAI and at Wiz, owned by Google, independently found the same flaws using AI tools.
The mechanism is less about AI discovering entirely new bug classes than about making a previously impractical attack practical. Wiz researcher Sagi Tzadik said the vulnerability had been known to the PHP team for years but was left unaddressed because it could not be leveraged by a human attacker. Advanced models helped analyze code, test exploitation paths, and validate which theoretical weaknesses could actually be triggered. Many exploits circulating publicly now appear ineffective, which underlines that a working chain still requires careful verification.
The incident fits a wider overload in vulnerability management rather than an isolated PHP problem. Microsoft security updates covered a record 1,453 vulnerabilities in a single month, increasing pressure on developers, administrators, and under-resourced IT teams. Bank of America recently hired 1,000 consultants to fix flaws found by Anthropic Mythos. Researchers stress that AI magnifies existing discovery and disclosure dynamics instead of creating a fundamentally new process, while coordinated patching remains the main defense.
What faster AI-assisted exploitation means for companies
For companies running web applications, the practical change is a shorter window between disclosure and mass scanning. PHP is widely exposed to the internet by design, so even an unconfirmed exploit can trigger millions of probes within hours. Larger organizations need automated asset inventory and patch deployment to cover PHP versions across sites, APIs, and legacy projects. Smaller firms with one administrator or an external contractor face higher risk from delay, because public scanning does not distinguish company size.
The limits of the current threat should shape purchasing and operational decisions. No mass takeovers or breaches have been confirmed, and the PHP Foundation says it has seen no successful exploitation. Frontier models still do not find all bugs from simple prompts, as Calif researcher Ryan Austin noted, and sophisticated cases combine AI with expert review. Buyers should ask vendors how PHP versions are tracked, how quickly patches are tested and rolled out, and what compensating controls exist when immediate patching is impossible.
The marker to watch is whether other long-lived flaws in widely deployed internet-facing software become practically exploitable after years of inactivity. Researcher Niels Provos doubts a flood of such cases, noting that AI-assisted review has run for over a year and PHP2Shell still required a highly skilled team. Former NSA hacker Jake Williams adds that few platforms match PHP reach. If similar historic bugs stay theoretical, current patching routines hold; if more turn operational, disclosure-to-patch timelines will need to shrink.
