When the PHP2Shell vulnerabilities were disclosed, security teams and web administrators rushed to apply patches while criminal actors launched attacks. Cloudflare’s telemetry recorded millions of exploitation attempts within hours of the public release. Analysts linked the speed of these attempts to the widespread availability of generative-AI tools that can automate the crafting of exploit code.
Despite the flood of attempts, the feared scenario of large-scale victim takeovers has not materialized. The PHP Foundation reported no confirmed successful compromises and noted that many of the publicly shared exploit scripts failed to achieve their intended effect. This suggests that, while automated tools can generate attack vectors quickly, the underlying code may still resist practical exploitation.
The first public identification of the PHP2Shell flaws came from Calif, a security firm that combined frontier AI models with deep knowledge of PHP internals. Working under OpenAI’s Patch the Planet initiative, Calif’s researchers produced functional exploits in early August, leveraging AI to navigate the complex code paths that had previously been deemed too intricate for manual exploitation.
Following Calif’s breakthrough, teams at OpenAI and the Google-owned security company Wiz independently uncovered additional PHP2Shell attack surfaces using similar AI-driven techniques. Wiz’s security researcher Sagi Tzadik explained that the vulnerability had been known to the PHP core team for years but remained unaddressed because it could not be weaponized by a human attacker, a limitation that AI tools have now begun to overcome.
The episode highlights how AI accelerates both discovery and exploitation of software flaws, intensifying pressure on under-funded IT departments and developers. In the same month, Microsoft released a record 1,453 security patches, while Bank of America contracted a thousand consultants to remediate vulnerabilities identified by Anthropic’s Mythos system. Researchers caution that AI is not inventing new attack methods but amplifying existing cycles of disclosure and response.