A security researcher used GPT-5.6 Sol to identify a PHP exploit chain before attackers found it, OpenAI said, and the company released a tool that lets developers check whether their servers are exposed.
The researcher worked with Calif, a partner in OpenAI’s Patch the Planet programme, during an audit funded by the lab in August. Patch the Planet pairs frontier models with human review on open-source code.
Ami Luttwak, co-founder of Google’s Wiz security team, told the Wall Street Journal the flaw ranks as the most severe AI-discovered vulnerability on record, and said it was probably too complex to have been found without AI. Wiz also helped trace the PHP2Shell exploit chain.
Why PHP is the test case
The language runs a large share of the public web, often on shared hosting where administrators update slowly or not at all. Reporting around the disclosure put the affected footprint at a large majority of sites.
A defect in a common component can sit unpatched for months while a fix exists upstream. OpenAI’s checker answers the first question quickly: is this server running a version that needs attention. It patches nothing.
The agent problem in reverse
The disclosure lands in the same week that agents were used as attack tools. Reuters reported that a hacker directed agentic software against more than 100 company sites for a few thousand dollars.
The same capability that helps a researcher enumerate flaws helps an intruder enumerate targets. Whether faster discovery reduces risk depends on whether defenders patch at the speed models can look.
