GPT-5.6 Finds WordPress Zero-Day RCE for $25 in API Costs, Raising Fresh Alarm Over AI in Hacking

Security researchers at Searchlight Cyber have demonstrated that OpenAI's GPT-5.6 coding model can uncover a serious remote code execution vulnerability in WordPress for approximately $25 in API costs, a finding that crystallizes growing concern over how frontier AI tools can be weaponized to find exploitable flaws at a fraction of their black-market value.

The research, published on Searchlight Cyber's website, shows that a single researcher, armed with GPT-5.6 Sol Ultra and a deliberately vulnerable WordPress codebase, was able to identify a chain leading from pre-authentication access all the way to remote code execution. The result lands at the center of a brewing debate about the offensive security potential of large language models, a theme tracked across latest AI developments.

A $25 discovery in a $500,000 market

The headline of the research is its stark economics. Exploit brokers, the intermediaries who buy vulnerabilities and resell them to governments and private buyers, will pay up to $500,000 for a reliable WordPress remote code execution flaw, according to Searchlight Cyber. WordPress powers roughly 40 percent of all websites, making a core RCE bug extraordinarily valuable.

Against that benchmark, the roughly $25 the researcher spent on GPT-5.6 API calls to find the flaw is arresting. The work did not require a large team, specialized fuzzing infrastructure, or weeks of manual auditing. Instead, it relied on directing a capable coding model to reason about WordPress's source code from first principles.

The researchers released a free checker tool, wp2Shell.com, so that administrators can determine whether their WordPress instances are affected. They said they held back full publication details to give defenders time to respond.

How the model found the flaw

At the heart of the discovery is WordPress's hook system, the mechanism that gives the platform its enormous plugin ecosystem. Hooks, with names like `wp_enqueue_scripts` or `publish_post`, allow plugins to attach custom behavior to nearly every stage of the WordPress lifecycle. They are divided into actions, which execute side effects, and filters, which modify data as it passes through.

That flexibility is also a sprawling attack surface. The task the researchers set was to find a chain that allows remote code execution from an unauthenticated starting point, with success defined as the ability to read a protected file from the server's filesystem.

GPT-5.6 Sol Ultra, OpenAI's reasoning-focused coding model, worked through the analysis by examining how data flows through WordPress's hooks and into potentially dangerous operations. The model's strength lay not in blindly searching for patterns, but in reasoning about the interactions between components, a capability that has improved markedly with recent reasoning-focused model releases.

The researchers framed the exercise around a repository containing a vulnerable copy of WordPress, a setup resembling a capture-the-flag challenge in which success meant extracting a flag file. But they stressed that the underlying vulnerability class is real and affects production deployments running MySQL, the database behind most live WordPress sites.

The dual-use dilemma

The finding lands as AI companies face intensifying scrutiny over how their models can be misused. Frontier models are increasingly evaluated for their ability to assist with cyber offense, and the results have grown more uncomfortable as capabilities climb. A model that can shave the cost of finding a critical bug from hundreds of thousands of dollars to the price of a few coffees reshapes the threat landscape in concrete ways.

Defenders argue that the same capability accelerates defensive work, helping security teams find and patch flaws before attackers do. Searchlight Cyber itself occupies both sides of that line, offering attack surface management and dark web monitoring tools while publishing offensive research. The wp2Shell checker embodies the defensive case: turn the discovery into something administrators can act on immediately.

But the asymmetry is hard to ignore. A lone researcher with an API key now has access to auditing power that once required a well-funded team. As the cost of finding vulnerabilities falls, the burden shifts to the speed and scale at which they can be patched.

What it means for WordPress and beyond

For the WordPress ecosystem, the research is a reminder that the platform's hook architecture, the foundation of its plugin economy, is also its most consequential weak point. Site operators are encouraged to run the wp2Shell checker, keep plugins and core updated, and restrict what unauthenticated users can reach.

More broadly, the experiment is a data point in a fast-moving debate about where the line should sit between open security research and the proliferation of offensive AI capability. As coding models grow more capable, demonstrations like this one are likely to become more common, not less.

Stay Ahead of AI

The intersection of AI and cybersecurity is moving faster than any single team can track. For reporting on offensive AI research, defensive tools, and the policy debates that follow, follow the AI industry coverage at AI Buzz Wire.

Read more AI news →