NSA and FBI urge fixes for a 14-year-old PHP flaw that OpenAI says its model helped find

The National Security Agency warned organisations to patch security flaws that were identified with the help of artificial intelligence¹. The NSA and the FBI have urged fixes for the flaws, which sit in widely used software⁶.

Cybersecurity researchers using AI found a bug in the PHP programming language, and the discovery has thrown parts of the internet into disarray^4,5. OpenAI said its GPT-5.6-Sol model helped find the flaws³.

The bug had been present in the code for 14 years, OpenAI President Greg Brockman said on Wednesday¹. Brockman said hackers could exploit the vulnerability to steal private user data or modify websites⁴.

By one estimate, the flaw could take down 70% of the internet². Brockman noted that PHP was his “first programming language”².

Why it matters to investors

The episode shows AI working on both sides of cybersecurity. A model found a flaw that had gone unnoticed for 14 years, and the same capability raises the question of who else can find such flaws first^1,3.

New research points the same way. The most advanced AI models are now capable of blueprinting cyberattacks on power grids and water systems with alarming success rates, according to cybersecurity company Frenos⁷. “It’s a clock that is moving faster than we’ve ever seen,” said Steve Swick, a former top security executive at American Electric Power⁷.

The researchers argue that critical infrastructure operators are running out of time, and that it is only a matter of time before freely available models make such attacks easier to carry out⁷.

Defenders face a separate problem. Security teams that rely on AI models to investigate attacks have found that the models can suddenly refuse to help, as strict guardrails lock them out mid-investigation⁸.

What to watch

The immediate task is patching. The official warnings put the onus on organisations running the affected PHP code to apply fixes^1,6.

AI tools are themselves targets. OpenCode, an open-source AI coding agent, said it was compromised over the weekend, allowing attackers to obtain API keys belonging to some customers, including keys for OpenAI, Anthropic and Google services⁸.

For companies, the combination of AI-found vulnerabilities, AI-planned attacks on infrastructure and compromised AI tools widens the range of cyber risk they face, at a time when defenders’ own AI tools can lock them out^4,7,8.