Inside Proton's AI Push: What 404 Media Reports, and What Is Confirmed

A new 404 Media investigation describes a company at war with itself over AI. Proton, the Swiss privacy company behind Proton Mail, Drive, Pass and the Lumo chatbot, is pressing its engineers to use AI far more, according to nine current and former employees, internal Slack messages, meeting audio and an internal document from the chief technology officer. 404 Media says Proton did not dispute the reporting.

This is a single-outlet story built on leaked material, so we attribute every claim and mark what we could and could not confirm. It is worth covering because it is a vivid case of a trend we have tracked across the industry: companies setting AI-use expectations for engineers, as with Amazon's engineer role guidelines and Uber putting ChatGPT in front of all 30,000 employees, but at a company whose product is trust.

TL;DR: what is reported, and by whom

What does the report say happened?

According to 404 Media's account, the story runs through 2026. In May, Proton launched post-quantum encryption that initially did not cover older emails, and the CEO objected to the public attention that limitation drew. In June it launched Lumo 2.0. In August, in a WIRED interview, Yen said Proton was building privacy-first AI so that privacy companies do not cede the future to Big Tech.

In late August, per the report, Yen told an employee in a security channel who wanted code penetration-tested before release that their view was "bullshit and loser thinking," urged engineers to use AI more, and said it would be the last year Proton hires engineers. On August 27 the company had a six-hour outage, which an internal document later cited as evidence of unreliable engineering. In September, CTO Bart Butler posted an "AI manifesto" to Slack that employees criticized as alarmist and exaggerated. In early October Butler reportedly said about 90% of Proton's code was written by AI, and employees say much of it uses Anthropic's Claude.

Most recently, Proton announced it would re-interview staff about their AI use. An email called the interviews "consequential," and employees believe they could affect whether they keep their jobs, according to 404 Media. A Tom's Guide report covers the same re-interview plan under the headline that jobs may be on the line; we could only read its headline, so we rely on 404 Media for the details.

Who said what

• Andy Yen (CEO): quoted telling employees to "stop lovingly handcrafting all our code and use AI more," and "The train is coming, get onboard or get crushed."
• Bart Butler (CTO): quoted as saying "It's because we ran out of time" about why the push is happening, and, in a meeting about the re-interviews, that "there's nothing like a deadline to motivate people."
• Employees: describe a culture of fear. One said, "A fear based culture feels like a cancer to me."
• Proton spokesperson: said some people are "unhappy with our direction, or misunderstand it," and that the company has "sometimes been too blunt or too slow in explaining ourselves."

These are quotations as reported by 404 Media from messages and recordings it says it reviewed. We have not seen the underlying material.

The internal document

404 Media says an internal CTO document states Proton is not on track to meet its annual recurring revenue target due to "insufficient velocity," describes the company as "an underperforming company" by key metrics, and proposes an "elite fast team" of 5 to 10% of staff with a 30% base salary bonus and weekend work. It reportedly considered moving headquarters to the US or China and rejected both, citing Switzerland's privacy protections. That context matters: the AI push is framed internally as a speed and growth response, not only a strategy about technology.

Why this matters beyond Proton

Security review is the heart of the dispute. The most specific allegation is that an employee asking for a penetration test before release was dismissed. If AI really writes most of the code, the volume of code needing review grows while headcount shrinks. Independent research already suggests the risk is real: see our coverage of the LAST secure-code benchmark, where vibe-coding agents produced insecure code in a large share of cases and the practical playbook for shipping AI-written code you do not understand.

For a privacy product, the stakes are higher than average. Customers choose Proton because it promises end-to-end encryption and careful engineering. Nothing in the report alleges a breach or a specific vulnerability, and we found none. It describes process and culture. But a company that sells trust is judged on process.

There is also a contrast with Yen's public line. In an August WIRED interview, as quoted by 404 Media, he said Proton builds AI tools because privacy companies that do not "build privacy-first AI" are "essentially ceding the future to tech companies that don't have the best interests of users at heart." The internal messages the report describes are far more aggressive in tone than that public framing.

An open briefcase with a notebook and sparkle, picturing Proton staff facing AI-use expectations at work

How does this compare with other AI-at-work mandates?

Most large-company AI rollouts we have covered are framed as enablement: tools, training and usage dashboards. Amazon's guidelines describe a "renaissance developer" who uses AI heavily, and Uber's pilot started with a thousand people before expanding. Even Anthropic, a lab that builds these models, has a manager arguing that hunger beats tenure in hiring, which is about attitude rather than tooling. What stands out in the Proton account is the combination of three things at once: a hiring freeze for engineers, an interview process that tests AI use, and an alleged dismissal of a request for security testing.

The key distinction is between measuring adoption and measuring outcomes. Counting how much code an assistant wrote tells you about tool usage; it says nothing about defect rates, incident counts or time to fix. A fair evaluation of an AI push would track escaped bugs, security findings per release and outage minutes before and after. If the August 27 outage was evidence of weak engineering, as the internal document reportedly argues, the same metrics should show whether the AI push helps.

Questions worth asking any software vendor

Whether or not every detail in this report holds up, it suggests a checklist for customers of security-sensitive software:

1. Who reviews AI-generated changes, and is review mandatory before release?
2. Are independent audits or penetration tests recent, and do they cover code written after the AI adoption?
3. How is the share of AI-written code defined and measured?
4. Are staff free to escalate security objections without penalty?
5. What happened in the last outage, and what changed afterward?

Vendors that answer these plainly deserve more trust than those that cite a percentage of AI-written code as a badge.

What is not confirmed

• The 90% AI-written code number has no published methodology; "written by AI" could mean anything from autocomplete to agent-authored pull requests.
• We have not seen Slack messages, audio or the CTO document; they are described by 404 Media.
• We could not find a Proton blog post or filing addressing the report in detail beyond the spokesperson statement.
• No outage cause, vulnerability or breach is linked to AI-written code in the reporting.

What this means for what you build or pay

• If you run an engineering team: an AI-use mandate without a review budget is the failure mode here. Set measurable expectations about review and testing, not only about usage.
• If you buy privacy software: ask vendors how AI-assisted code is reviewed and whether independent audits cover recent releases. For AI agents that touch your systems, a control layer such as AgentBeam, from the explainx.ai team, can block dangerous agent actions before they run.
• If you are an engineer: document your review decisions and push for security sign-off to be a stated gate, not an informal one.
• If you are an analyst: treat percentage-of-code-written-by-AI claims as marketing until a method is stated.

Bottom line

The report is detailed, mostly sourced to named internal channels and a company that did not dispute it, but it is one outlet's account of leaked material. Its enduring value is as a case study: AI adoption pressure, security review and layoffs fear collide most visibly at companies whose brand is trust. We will update this article if Proton publishes its own account or further outlets confirm details.

Related reading

• Amazon's software engineer role guideline and AI token metrics
• Uber gives all 30,000 employees ChatGPT
• LAST secure-code benchmark for vibe-coding agents
• Shipping AI-written code you do not understand
• Brave Leo end-to-end encrypted AI chat
• Anthropic manager drops coding interviews
• Official: 404 Media report and Proton's Lumo announcement

Details are accurate as of October 29, 2026 and rely on 404 Media's reporting.