The Pulse
AI Helped Turn Old PHP Bugs Into a Race to Patch the Web
Millions of attempted attacks hit PHP flaws within hours of their public disclosure on October 21, 2026, though the PHP Foundation said it had seen no successful exploitation. AI tools helped researchers turn long-known weaknesses into prac
AI.info Team · Written with AI · fact-checked ·
Millions of attempted exploits hit PHP systems within hours of the flaws’ public disclosure on October 21, according to Cloudflare analysis cited by WIRED. The PHP Foundation said it had not seen successful exploitation of the vulnerabilities, known together as PHP2Shell. The gap between those two facts captures the new urgency: AI helped researchers establish how to exploit weaknesses that had been known, in part, for years, while attackers moved quickly once details became public.
Old weaknesses, newly practical attacks
The flaws affect PHP, the scripting language used by a large share of websites. Researchers at Calif found the vulnerabilities in August while reviewing PHP with AI tools and their own knowledge of the language. OpenAI says Calif used GPT-5.6 Sol through its Patch the Planet initiative; researchers at Wiz independently found and reported both flaws.
Some of the underlying issues surfaced during a 2024 PHP security audit, but researchers had not then demonstrated a workable exploit. Calif’s account traces one related out-of-memory problem to an earlier report and describes how the team used an AI assistant to investigate PHP’s memory handling and identify paths to remote code execution. The discovery did not come from a model searching without direction: Calif says its researcher drew on PHP’s vulnerability history and asked targeted questions.
AI shortened the path from bug to exploit
The practical change was not simply spotting suspicious code. Calif’s write-up describes researchers using AI to help reason through how memory exhaustion and interrupted cleanup could leave PHP in a vulnerable state, then extending the pattern to additional paths. The firm says the resulting attack paths affected different PHP configurations, including Apache’s mod_php and PHP-FPM, with some exploitability depending on the application and setup.
“The vulnerability had been known to the PHP team for years, but because it could not be leveraged by an attacker—at least, not a human one—it was not addressed,” Sagi Tzadik, a security researcher at Wiz, told WIRED. His account points to a shift in what counts as an exploitable flaw: a weakness that once appeared too difficult to turn into an attack may become more urgent when AI can help researchers work through the steps.
Calif researcher Ryan Austin cautioned against treating the work as an autonomous discovery. “No model is simply capable of unearthing ‘all bugs’ with simple prompts. At least for now,” he wrote in the firm’s account of the research. Human expertise, target-specific knowledge and verification remained part of the process.
The disclosure window narrowed
OpenAI says it coordinated with PHP maintainers, researchers and security providers to test fixes and prepare protections before public disclosure. The company wrote, “To our knowledge, frontier AI found these vulnerabilities before attackers did,” describing the period available to prepare patches as a “Defender’s Window.” That is OpenAI’s assessment, not proof that no attacker had found the bugs independently.
Once the vulnerabilities became public, attempted exploitation followed quickly. WIRED reported that the PHP Foundation had not identified successful attacks, while the Foundation’s own account described the effort to prepare and release fixes amid growing public attention. A rapid increase in discovery can give defenders earlier warning, but it also leaves maintainers and operators less time to coordinate before attackers begin testing exposed systems.
Maintainers and operators face the backlog
PHP’s episode puts pressure on both sides of the response. Maintainers must assess reports, settle on fixes that do not break compatibility and prepare releases across supported versions; organizations then have to identify which servers run affected PHP and apply updates. OpenAI’s proposed response framework includes involving distributors and hosting providers before disclosure so they can prepare updates and interim protections.
The immediate task is patching systems that run vulnerable PHP, not assuming that a lack of confirmed compromises means there is no exposure. The Foundation said its security process had struggled with hard trade-offs around the earlier issue, while the AI-assisted research made the exploit path clearer and moved the fix to the top of its priorities. How many internet-facing systems remain on vulnerable or unsupported releases—and how quickly their operators can update them—now determines how long the attackers’ window stays open.