AI-Assisted Research Turned Years-Old PHP Bugs Into an Active Exploit Wave

AI-Assisted Discovery Surfaced Critical PHP Vulnerabilities

Security researchers at Calif used frontier AI models to uncover PHP2Shell, a set of long-dormant vulnerabilities in the PHP web scripting language that attackers began probing within hours of public disclosure. According to reporting from Wired, Cloudflare logged millions of exploitation attempts in that window, even though the PHP Foundation has not confirmed any successful breaches so far. The episode illustrates how AI-assisted vulnerability research can accelerate the discovery and weaponization of bugs that were previously thought too complex for human attackers to exploit at scale.

Discovery Timeline and Involvement of Frontier AI Models

Calif researchers first identified the PHP2Shell flaws in early August, partly through frontier AI tools used as part of OpenAI's Patch the Planet program and partly through expert human knowledge of PHP's internals. Some of the underlying issues were flagged as far back as 2024 but were considered too complex for a human attacker to practically exploit at the time. The discovery marked a turning point: what security teams had acknowledged for years but deprioritized suddenly became an active threat. After Calif's discovery, researchers at OpenAI and the Google-owned security firm Wiz independently rediscovered PHP2Shell using AI tools of their own.

According to Wiz researcher Sagi Tzadik, 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. This reasoning reflects the conventional prioritization model in software maintenance: bugs that cannot be practically weaponized can be deferred in favor of more immediately exploitable flaws. Frontier AI models disrupted that calculus by making previously impractical attack vectors suddenly feasible.

Human Expertise Remained Essential to Weaponization

Researchers emphasized that AI's role in uncovering PHP2Shell was catalytic, not autonomous. Ryan Austin, the Calif researcher who first surfaced the flaws, wrote that no current model can simply unearth every bug with a casual prompt, and that human expertise remained essential to turning the research into working exploits. The discovery process combined frontier AI's pattern-matching capabilities with deep domain knowledge—the kind acquired over years of working with PHP's internal architecture, dependency chains, and edge-case behavior.

This hybrid approach mirrored findings from security researchers across the industry. Jake Williams, vice president of research and development at Hunter Strategy and a former NSA hacker, told Wired that PHP's broad internet exposure makes it unusually fertile ground for this kind of AI-assisted rediscovery, though he does not expect an endless supply of similarly exploitable legacy bugs. Williams's framing captures an important constraint: not all software sits at the intersection of ubiquity, legacy complexity, and internet-facing deployment that makes AI-assisted rediscovery practical. PHP's position in web infrastructure—powering an estimated 77% of all websites with a known server technology—makes it a uniquely attractive target.

Speed of Attacker Response and Defensive Workload

The velocity of the attack wave exposed defenders to compressed decision timelines. Millions of exploitation attempts hit within hours of the vulnerabilities being published, according to Cloudflare's telemetry. However, many publicly circulating potential exploits appear to be ineffective, and the PHP Foundation reports no confirmed successful breaches. This gap between attempt volume and successful compromise suggests that publicly available, AI-generated exploit code may contain errors or rely on assumptions about target infrastructure that do not hold across diverse deployments.

The defensive response to PHP2Shell contributed to a broader surge in patch workload. Microsoft shipped a record 1,453 security fixes this month, and Bank of America reportedly hired 1,000 consultants to patch flaws surfaced by Anthropic's Mythos tool. These numbers underscore the cascading effect of AI-assisted vulnerability discovery: security teams and organizations must now resource patch management at unprecedented scale, even for tools that may not yet have demonstrated real-world compromise.

Threat Landscape Assessment and Historical Context

Security researchers emphasize that AI is ultimately not doing anything new, but rather magnifying existing dynamics of vulnerability discovery and coordinated disclosure. Vulnerability research has always benefited from automation and tooling; AI represents an incremental but meaningful step forward in that trajectory. We will definitely see more of these long-life bugs that are in every version of the software for the last decade or whatever, says Williams, but there is only so much of this low-hanging fruit. There isn't much else that has the specific reach of PHP. It's a pretty wild attack surface—most other tools aren't exposed to the internet by design.

Security engineer and researcher Niels Provos told Wired he isn't alarmed yet, pointing out that AI-assisted bug hunting has been underway for more than a year and still required a highly skilled team to crack PHP2Shell. Provos's position reflects a measured view of AI's scaling properties: even with advanced models, moving from theoretical vulnerability to production-ready exploit demands both technical depth and time investment. The PHP2Shell case required researchers to understand not just the vulnerability itself, but how to reliably trigger it, chain it with other flaws, and bypass common mitigations—work that frontier AI assists but does not fully automate.

Implications for Software Maintenance and Industry Practice

PHP2Shell represents a concrete validation of a long-standing security principle: the longer a flaw goes unpatched, the more dangerous it becomes once discovered. The incident demonstrates that "patient" threats—bugs known internally but deprioritized—now carry elevated risk as AI tools lower the barrier to weaponization. Organizations managing legacy systems with years of accumulated technical debt face an asymmetric problem: defenders must maintain patch velocity proportional to the fastest attacker equipped with frontier models, while maintaining the systems integrity and stability that operations depends on.

The economics of this shift affect security budgeting and staffing. If AI-assisted rediscovery means that patching workloads spike unpredictably, security teams must either maintain surge capacity or accept heightened risk during processing delays. Bank of America's decision to hire 1,000 external consultants illustrates one response; smaller organizations lack that economic flexibility and will likely face harder trade-offs between patch speed and other operational priorities.