For the past year, the conversation about enterprise AI and security has fixated on what region the models come from. It is a fair question. But as the cofounder of the world’s largest platform to run open models locally, I often speak with the increasing number of business leaders adopting open models. Where their models come from is not the question that keeps them up at night.
What they worry about is more simple. They ask questions like, “Where does my data go when I use these models?”, “Who can see it?”, and “How long is it kept?”
These questions matter more than ever right now, and not just because of the latest string of hacks over the past month, including PHP2Shell which left millions of websites around the world vulnerable, but because of how AI is being used in businesses. Coding agents, security tools and internal assistants send enormous amounts of company information to a model: source code, internal documents, customer details, the context of a team’s daily work. Not only is this highly sensitive information for a business, but it’s also extremely valuable to anyone who wants to train a model of their own.
AI leaders are already bracing for what comes next. NVIDIA’s Jensen Huang made it a feature in his keynote at GTC Berlin, and Jason Clinton, Anthropic’s deputy chief information security officer, recently warned that conditions would get worse before they improve, urging business leaders to take action. That action starts with knowing where company information goes when AI is used.
Most companies assume a vendor either protects their data or it doesn’t. The reality is not as black-and-white.
There are cheap, readily available services that openly trade lower prices for the right to train on the data you send. That can be a reasonable deal, as long as the customer knows they are making it. There are also providers that promise zero data retention: prompts and responses are processed, not stored, and never used for training. In between those two, sits a large gray area of providers that promise not to train on your data but may keep it for a number of days.
None of these are inherently wrong. But in a world where the portion of developers that use coding agents seems to be heading toward 100%, enterprise leaders need to understand the risks of an agent acting on a developer’s behalf. An agent can end up on the wrong end of that provider spectrum, without any developer actually deciding to put them there. As demand for tokens has surged, providers have taken on more traffic. Not all of them have built the controls to match.
Enterprises should treat data retention the way they treat any other vendor risk: ask the question directly, get the answer in writing, and make zero retention the default.
Then there’s the question about models. American or Chinese? Closed or open? Open models have historically been framed as less secure, but in practice, many enterprises find the opposite. Companies can run them in their own environment, so data never leaves. Or they can run them with a provider they trust, in a region their compliance teams have approved. Either way, the company chooses where its data goes.
Menlo Ventures’ latest data on the state of AI in the enterprise reported that open-source models’ share of usage is on the rise, suggesting significant future growth. About 10 million developers around the world use Ollama every month to run open models, processing trillions of tokens per day. Eighty-five percent of Fortune 500 companies have Ollama installed on desktops to run open models locally. They’re integrating open models into their workflows for a variety of reasons, among them open models’ lower costs and superior or equal performance on most software development tasks. On coding tasks, for example, the open source model DeepSeek V4 Flash scores within 0.6 points of Anthropic’s Opus 5, at about 1/13 of the cost.
A fintech company we work with told us that they’d burned 60% of their annual budget three months into a contract with a frontier model. Eighty-five percent of that spend came from engineers defaulting to the most expensive models, leading them to test open models for about a year. On code review, an open model (Deepseek) was faster and cheaper than Anthropic’s Sonnet. Across routine engineering work, open models performed essentially the same as Anthropic models. After two weeks routing most engineering work to open models, while still keeping frontier models for complex problems, they saw an immediate 30% savings on token costs.
Another one of our customers, a large consumer AI company, was spending more than $10 million on frontier models. In their own (human) evaluation, an open model (Gemma 4) matched the frontier model they were using at roughly 1/10th of the cost.
Open models are also becoming a practical tool for defense. Security teams need to work closely with the techniques attackers use: analyzing malicious code, probing their own systems, responding to an incident in real time. Access to that kind of capability on closed models can be limited to specific programs controlled by the labs that build them, where open models are available to any team that needs them.
Of course, open models are not immune to misuse, but the answer isn’t to leave defenders without options. It is to run them carefully, on infrastructure you trust, with clear rules for your data. Even the labs say so. Anthropic’s S-1 filing acknowledges that cyber and safety risks may delay the release of its most capable models. That can be a responsible choice, but it means defenders who depend on a single gated provider can find themselves without access when they need it most.
Models have advanced faster than most corporate AI policies. Many companies that were wary of open models a year ago are now comfortable with them, including models developed outside the United States, provided they run on US or European infrastructure under strict data terms. Their policies should catch up to that reality. Three questions are a good place to start:
Where does the model run, and who else touches the data? Know every party in the chain.What is the retention policy, and is it zero? Get a clear answer, and ask how it is enforced.Does your security team have the tools it needs? If frontier tools are gated, make sure there is a second option.
AI is becoming part of nearly every workflow in the enterprise. The companies that handle it well will not be the ones that picked the “right” model. They will be the ones that stayed in control of their data.The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
Click Here For The Original Source.
——————————————————–
