GPT-5.6 Sol Helped Identify Two Critical Vulnerabilities in PHP

Two flaws in PHP’s core engine could let a remote attacker run code on an affected server or expose secrets stored in its memory. OpenAI says researchers using GPT-5.6 Sol helped identify how to exploit the weaknesses, which had been known to the PHP Foundation since 2024. The episode shows how AI tools can help turn a suspected defect into a practical security finding—and how quickly maintainers may need to respond once that happens.

What the flaws could expose

The vulnerabilities, recorded as CVE-2026-109618 and CVE-2026-109619, concern PHP’s handling of memory and errors. Both received a severity score of 9.2 out of 10. Because the defects sit in PHP’s engine, a website could be affected even if its own application code contains no security flaw.

The first issue can arise when PHP runs out of allocated memory while handling a request. That failure may leave incomplete internal data behind. If the engine later treats that data as valid, an attacker could potentially exploit the resulting state. The second flaw involves a fatal error that leaves PHP referring to memory that has already been freed. A later request may then access or alter that memory.

In certain conditions, the source says, either weakness could be exploited remotely without login credentials to execute code on a vulnerable server. Researchers also demonstrated that the flaws could expose information held in server memory, including passwords, authentication tokens and API keys. The practical impact depends on the server and the conditions needed to trigger an exploit, but the combination of remote code execution and possible secret exposure makes these engine-level issues significant for PHP operators.

Fixes are available for supported versions

The PHP Foundation issued patched releases PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12. PHP 8.1 and earlier versions will not receive official fixes, according to the source. Sites running those older releases therefore face a different maintenance decision from sites that can move to a patched branch.

The coordinated release was planned for Oct. 22, then moved up to Oct. 21 after information about the flaws became public on Oct. 20. The two-month repair process involved PHP maintainers, researchers and security providers. OpenAI attributed the schedule to the complexity of the work, which required substantial changes to the PHP engine.

OpenAI says sites using PHP can upgrade to a patched version. For operators, the first practical step is to identify which PHP release their servers run and whether it is covered by one of the fixes. The source does not provide details about specific deployment procedures or the effect of upgrading on individual applications; teams will need to assess those against their own environments.

From a known defect to a demonstrated exploit

Researchers from Calif, a California-based cybersecurity firm and founding participant in OpenAI’s Patch the Planet program, began auditing PHP with OpenAI tools on July 30, 2026, at the PHP Foundation’s request. The Foundation had known about the underlying flaw since 2024 but had not found a way to exploit it or a fix.

Calif researcher Ryan A. identified an exploitable vulnerability on Aug. 11, and OpenAI researchers did so the same day. Wiz later found another vulnerability independently. That sequence matters: AI assistance contributed to finding a working exploitation path, while a separate firm also identified a flaw. The source does not specify which individual research steps were performed by the model or how much time the tools saved.

OpenAI frames the case through its “Defender’s Window” concept: a limited period in which organizations can use advanced AI systems to discover and address weaknesses before similar capabilities become more widely available to attackers. The company says human developers and security teams had examined the code without fully resolving the issue, while AI-assisted researchers demonstrated practical exploitation methods within two weeks.

There is another interpretation. If no one had found a usable exploit, the long-known defect might have remained difficult to abuse in practice. But that uncertainty cuts both ways: once an exploit path is demonstrated, maintainers and operators have a more concrete problem to address. The useful question for defenders is how quickly they can move from a research finding to a verified fix across the systems they run.

Why PHP’s reach matters

PHP runs an estimated 70% of the internet, according to the source. That figure describes the language’s reach, not the share of sites confirmed to be vulnerable. Exposure depends on factors such as the PHP version in use and whether the relevant conditions can be triggered. Still, a flaw in a widely deployed engine can affect sites whose teams did not introduce the defect in their own code.

The case also sits within a broader debate about AI-supported vulnerability research. The source points to Anthropic’s report that Claude Mythos Preview identified thousands of zero-day vulnerabilities across major operating systems and web browsers. These examples suggest that researchers may be able to use AI tools to investigate complex software at a pace that changes how quickly maintainers need to assess and repair findings. They do not, by themselves, establish how often such work will produce exploitable bugs or how reliably tools can identify them.

The immediate takeaway is concrete: PHP operators should check their versions and plan upgrades to the patched releases available for their supported branch. The longer-term test is whether AI-assisted discovery can be matched by careful validation, coordinated repair and timely deployment, especially for software that remains widely used across many independently managed sites.

• AI data centers and pressure on memory and storage supply
• Rising demand for data center CPUs and AI agents
• Chip supply pressures affecting automotive production
• Custom AI accelerator designs

OpenAI’s account of the coordinated vulnerability disclosure

Background on OpenAI’s Defender’s Window concept

Anthropic’s reported findings from Claude Mythos Preview