Proton CEO Andy Yen publicly expresses reservations about artificial intelligence, but internally pushes for a very rapid transition to AI-assisted software development. According to reported information, the company, whose public image is built on privacy protection and cybersecurity, has announced that 2026 will be the last year it hires so-called traditional developers.
Starting in 2027, Proton plans to hire exclusively product managers tasked with steering AI to generate applications. According to management, the developer’s role is no longer to write code by hand, but to write specifications and prompts for models like Anthropic’s Claude to do the work. In October 2026, the company’s CTO claimed that 90% of Proton’s code was already generated by artificial intelligence.
Now, the priority is speed. When a security employee suggested spending a few days testing (pentesting) AI-generated code before deployment, Andy Yen dismissed this approach as “bullshit and a loser mentality.” In a post on X, he wrote: “Stop always thinking about how to slow down, and focus more on how to go faster.”
Monitoring and Re-interviews
To enforce this direction, management has installed tracking software to measure the amount of AI-generated code produced by each employee. Furthermore, the company is conducting re-interviews with its teams to assess their AI proficiency, explicitly warning that the results will influence their job security. This stance contrasts with Proton’s public image and is causing significant internal concern about the future security of the company’s software.
The debate also centers on the trade-off between deployment speed and security. Andy Yen has asked his teams to stop “lovingly handcrafting all the code.” For software professionals, bypassing human security reviews on AI-generated code is considered a major technical heresy, especially given that Proton’s users place a high value on security and privacy. A user summarized this viewpoint on X: “Vibecoding your entire service conflicts with these values.”
Developers and customers also point out a contradiction. Proton has historically positioned itself as the ethical alternative to tech giants, but its new software strategy relies heavily on third-party proprietary large language models, notably Anthropic’s Claude. According to these critics, the company is thereby transferring its technological value to the infrastructures of the Big Tech firms it publicly opposes.
A Quality Paradox
A large-scale analysis by GitClear of over 211 million lines of code, including repositories from companies like Google, Microsoft, and Meta, reports a degradation in code structure since the widespread adoption of AI assistants. The rate of code block duplication has quadrupled, according to the publication. For the first time, copy-pasted code has surpassed moved or reused code.
A case study on secure engineering further indicates that 87% of AI-generated Pull Requests contained at least one exploitable vulnerability. Researchers noted SQL injection, XSS, buffer overflow, and hardcoded credential exposure flaws. According to the article, AI reproduces the statistical patterns of its training dataset, often riddled with old, insecure code, which can produce seemingly functional code that is fragile from a security standpoint.
Andy Yen’s Response
Andy Yen argues that artificial intelligence and security can coexist in software development. On X, he states that “companies do performance reviews, and AI is part of that review in most places,” and that his positions on AI are “far more nuanced” than the criticized article suggests.
He frequently cites Lumo, Proton’s AI assistant, designed to address the privacy shortcomings of tools like ChatGPT or Claude. Initially, Lumo did not protect conversations with end-to-end encryption during their execution, with Proton merely promising not to log the sessions. To address the limitations of this promise, Andy Yen announced a transition to a secure enclave model on Proton’s servers. The goal is to isolate the execution of the large language model in a hardware-encrypted environment, technically preventing Proton from intercepting data during processing.
