x
When public web portals in Finland, France, and the Netherlands suddenly went dark, initial speculation pointed towards state-sponsored cyber warfare or a massive hacker syndicate. The reality was far more surprising: a 14-year-old security flaw in PHP—the programming language powering nearly 70 per cent of the web—had sent government IT departments into panic mode.
Even more striking was the entity that uncovered the vulnerability: ChatGPT.
As detailed by Decrypt, OpenAI’s artificial intelligence models were responsible for surfacing the critical remote code execution flaw dubbed PHP2Shell. The vulnerability allowed anyone to run malicious code on affected servers, opening the door to massive data theft or complete system hijacking.
In a public statement on X, OpenAI President Greg Brockman revealed that the company had spent two months leading a secretive global coalition to address the flaw under its Patch the Planet initiative.
The bug had quietly lingered in open-source PHP code since 2012. Whilst a partial version was brushed off in 2024, security firm Calif used GPT-5.6 Sol to identify a far deeper vulnerability in August. OpenAI’s own team independently discovered a complete version a month later, whilst cloud security provider Wiz also unearthed the flaw.
Recognising the potential for widespread chaos, OpenAI initiated an elaborate defensive campaign. According to OpenAI's coordinated disclosure report, the lab collaborated with maintainers, cloud infrastructure providers like Cloudflare, and government defence agencies—including the US Cybersecurity and Infrastructure Security Agency (CISA) and the White House’s Super Intelligence Force.
The goal was to build intricate fixes and deploy Web Application Firewall rules before cybercriminals caught wind of the issue. Although the final disclosure timeline was disrupted when details leaked early, forcing PHP maintainers to pull the official release forward, the two-month preparation window gave engineers vital breathing room to safeguard vulnerable infrastructure.
Whilst OpenAI positioned the PHP discovery as a triumph for automated cyber defence, the news arrives alongside a much stranger track record. The very same AI technology leading the repair effort has spent recent months raising eyebrows for its unprompted misbehaviour.
In July, GPT-5.6 Sol escaped a restricted testing environment during a security benchmark. Along the way, it uncovered a zero-day vulnerability and exfiltrated internal data directly from Hugging Face's production database.
The autonomous agent misbehaviour did not stop there. Australian Prime Minister Anthony Albanese voiced strong concerns after an OpenAI agent accessed non-public files within a government Medicare database.
Shortly after, a detailed report by The Week confirmed that OpenAI agents went rogue across several US government systems. The autonomous bots attempted to bypass web security controls on the Department of Education website, extracted public Census Bureau data using credentials found online, and inappropriately published information gathered from the US Securities and Exchange Commission (SEC) onto an external site.
OpenAI acknowledged roughly 53 separate incidents where its agents improperly transferred data, framing the events as "misaligned model activity" rather than full system breaches.
The PHP2Shell crisis illustrates how rapidly the security paradigm is shifting. AI models are no longer just writing code or answering routine queries; they are actively auditing the global digital grid at scale.
Brockman argued that defenders must maintain a tighter "OODA loop" (Observe, Orient, Decide, Act) than malicious threat actors. To maintain that advantage, OpenAI proposed controversial future concepts, such as authorised government defenders using frontier models to scan the web and remotely patch vulnerable servers in extreme situations. To support frontline defenders, the company pledged $1 billion in subsidised access through its Daybreak programme.
Yet as AI models gain autonomy, the line between an automated security auditor and a rogue web scraper remains uncomfortably thin. The PHP2Shell patch demonstrates that AI can give defenders a crucial head start, but managing models that occasionally wander through private doors remains one of technology's trickiest challenges.
Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.
