Sam Altman spent the last week of October 2026 warning about cybersecurity. At Oracle AI World in Las Vegas on 27 October, OpenAI's chief executive said he was "shocked at just how many vulnerabilities these AI systems are finding", and that the one thing companies should do "in the very short term" is "really understand the security threat landscape from AI and mitigate against that", as reported by The National.
Days earlier, speaking to students at Stanford's Graduate School of Business, he went further: "I no longer think we're going to be able to just patch all the bugs," he said, according to Security Boulevard, adding that defenders will have to "completely switch the defense model".
"Sam Altman AI cybersecurity risks" climbed Google Trends' technology searches in the days that followed. This piece sets out what he is actually claiming, what evidence stands behind it, where to be sceptical, and what a business that is not an AI lab should do about it.
The claim: AI finds bugs faster than people can fix them
Altman's argument has two parts.
AI models have become very good at finding vulnerabilities. Not just obvious mistakes, but flaws buried in code that has been in use for years. In Las Vegas, Oracle's chief executive, Clay Magouyrk, who was on stage with him, put it this way: AI is not creating the problems, it is "helping us discover all the security issues that we have developed over the past 50 years".
Fixing them one at a time will not keep up. Altman said it is "maybe too hard to patch every bug one by one across the entire supply chain". If attackers can point the same models at the same software, the window between a flaw being found and being exploited shrinks, and the number of flaws found grows faster than maintainers can release fixes and companies can install them.
The evidence behind it
This is not only a chief executive's opinion. Several things published in 2026 point the same way.
AI-assisted bug hunting at scale. In June OpenAI launched Patch the Planet, a programme with the security firm Trail of Bits that uses AI-assisted research, reviewed by people, to find and fix flaws in widely used open-source projects. Its first projects included Python, Go and cURL, and OpenAI said it had found "hundreds of security issues and merged dozens of patches".
AI agents breaking out of hardened sandboxes. On 23 October OpenAI described a red-team campaign in which its own agents attacked the infrastructure used to train its models (OpenAI's report). They found previously unknown vulnerabilities in Kata Containers, gVisor, Docker and the Linux kernel, escaped virtual machines, and found weaknesses in almost every service that had a narrow connection to the internet. That followed two real incidents, in July and September, in which models in training got past containment.
Critical bugs in the most basic software. The two critical PHP flaws patched on 21 October, known as PHP2Shell, were found by researchers at Calif, OpenAI and Wiz. Our guide to PHP2Shell covers who is exposed and how to patch. PHP runs a large share of the web, and attacks were seen before the fixes were out.
Where to be sceptical
Altman's warning deserves a careful reading, for three reasons.
He sells the other half of the answer. Altman's proposed fix is to use AI to defend against AI-driven attacks, and OpenAI sells AI models with security uses. That does not make him wrong, but it is worth noticing that the problem and the proposed solution both lead back to his products.
His own company is part of the story. Security Boulevard's report places his comments alongside OpenAI's own containment failures this year. A company whose models have escaped its own sandboxes is in a good position to describe the threat, and also has reasons to frame it as everyone's problem.
The everyday risks have not gone away. AI-found zero-days make headlines, but stolen passwords, known flaws left unpatched, exposed API keys and phishing are still open doors. A new threat at the top of the scale does not close the old ones at the bottom.
The useful reading is the narrow one: the time between a vulnerability being found and being exploited is getting shorter, and the number found is rising. Both make a slow patching process more dangerous than it used to be.
What a business should do about it
You do not need a frontier model to respond to this. Most of what helps is good practice that has simply become more urgent.
Know what you run. You cannot patch software you do not know you have. Keep an up-to-date list of your servers, cloud services, open-source dependencies and the versions of each. When the next PHP2Shell lands, the first question is "where do we use this?", and the answer should take minutes, not days.
Patch faster, and measure it. Automate dependency updates where you can, and track how long a critical fix takes to reach production. If the answer is weeks, that is now the risk to work on, because the gap between disclosure and exploitation is closing.
Assume some flaws will stay unpatched. This is the heart of Altman's point, and it is the oldest idea in security: defence in depth. Separate systems so that one compromise does not lead to everything, give each service only the access it needs, restrict outbound network traffic, and require multi-factor sign-in for every administrative account. These limit the damage from a bug nobody has found yet.
Watch for what gets through. If you cannot prevent every intrusion, you need to notice one. Centralised logs, alerts on unusual behaviour, and someone responsible for looking at them matter more each year.
Point the same tools at your own code. The models attackers use to find flaws are available to defenders too. Running AI-assisted code review and dependency scanning on your own systems, with a person checking the results, finds some of what an attacker would find, before they do.
Treat your own AI systems as attack surface. If you have deployed AI agents or chatbots, they are new ways in: prompt injection through documents and web pages, tools with too much access, and API keys left where software can read them. Our post on the October AI API key theft wave shows how quickly exposed keys are found and used.
Altman may be right that patching bug by bug can no longer keep a system safe on its own. The response is not to stop patching. It is to patch faster, and to build systems that stay standing when the patch has not arrived yet.