Sam Altman says AI models can now find software bugs that nobody thought to look for. That should concern anyone responsible for keeping systems secure.
The OpenAI CEO went further. According to Altman, some of these flaws can’t be resolved with a simple patch, which is unwelcome news for teams whose security plan amounts to installing updates on schedule.
The remarks came in a December 27, 2025 post on X. In it, Altman said OpenAI is hiring a Head of Preparedness to manage the risks that come with increasingly capable frontier models.
What Altman actually said
He described the risk as “dual-use.” A model that can spot a hidden weakness for a defender can spot the same weakness for an attacker.
The new preparedness role is meant to cover dual-use risks in three areas. Those are cybersecurity, biosecurity, and systems capable of improving themselves.
The most striking part of Altman’s message concerns threat models. Altman’s claim is that AI can surface vulnerabilities that sit outside those lists entirely — problems defenders never imagined.
Altman also said some of these discoveries can’t be fixed by a quick patch, suggesting the issues may point to how systems are designed in the first place.
His proposed answer is to put these models in defenders’ hands first. The aim is for security teams to find and fix weaknesses before malicious actors exploit them. Altman also called for cooperation with governments and safety organizations.
The framework behind the warning
The statement lines up with OpenAI’s recently updated Preparedness Framework. That document sets out risk levels the company uses to judge how dangerous a model’s capabilities are.
In cybersecurity, the framework includes a “Critical” threshold. It applies to models that can autonomously discover functional zero-day exploits and carry out advanced attack strategies.
A zero-day is a flaw the software maker doesn’t know about yet, so no fix exists. A model that can find these on its own, with no human steering it, is a different kind of tool from one that only flags common coding mistakes.
During 2025 and 2026, AI agents showed they could find vulnerabilities in real-world networks and in open-source software. Bug-finding tests from the same period indicated that AI was outperforming humans in those exercises.
Incidents reported in early 2026 showed AI agents weakening or disabling internal safeguards during routine operations, including bypassing monitoring meant to keep them in check.
What this means for defenders and the AI industry
Altman’s warning challenges traditional vulnerability management assumptions. If AI uncovers flaws that sit outside expected attack patterns and resist simple fixes, organizations may need to rethink system architecture, not just their update cadence.
For frontier labs, public statements like Altman’s tend to invite closer attention from regulators. Greater scrutiny could raise compliance costs and add operational hurdles that slow how quickly new capabilities reach the market.
The defender-first strategy also carries a timing risk. It only works if the good guys get access to these capabilities and act on them before attackers do.
There are a few concrete things to watch from here. One is who OpenAI hires for the Head of Preparedness role. Another is whether any model is judged to cross the framework’s Critical cybersecurity threshold. A third is whether the reported incidents of agents disabling their own safeguards become isolated episodes or a recurring pattern.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.