Proton cancelled its AI re-interviews; one employee says the test is moving into performance reviews

404 Media reported that on Friday Proton's chief executive, Andy Yen, cancelled a series of near company-wide "re-interviews". In them, employees would have had to explain how they use AI. Six current employees and audio of a roughly two-hour all-hands meeting, both obtained by the outlet, made the purpose plain: staff judged to be using AI too little could be fired. The reversal came a day after 404 Media published an investigation into how Yen was pushing AI across Proton, in some cases over employees' security objections. The headline reads as a climbdown. What was said inside the all-hands shows the assessment moving into the performance review.

The test moved into the review cycle

Yen opened by acknowledging "a communication failure on my part". Then he said: "Going forward, I'm still going to push." He told staff the re-interviews had "already served their purpose" in an unintentional way, because employees were now talking about AI and learning it.

One employee told 404 Media that leadership said "in the next breath" they would keep pushing AI, and that the things the re-interviews were meant to measure "would become a component of the performance reviews." Another put it more bluntly: "They have said clearly the direction had not changed."

Yen also announced a reorganisation for next month, describing part of it as "redistributing talent across teams". He added that "there will be no loss of jobs from AI." The CTO, Bart Butler, had already set the terms in an AI manifesto posted to Slack, which told engineers: "Your old job is gone."

Put those three statements side by side. Proton staff now face an AI-use standard written into their reviews and a reorganisation that moves people between teams. Their leadership has also told them in writing that the role they were hired for has been redefined. The formal test is gone. The grading remains. If your employer has started weighing AI use, the guide on how to show impact when your employer grades your AI use sets out what to do before your next review.

The objection was judgement, and the metric cannot see it

The exchange that defined the story happened in Proton's security Slack channel at the end of August. An employee asked the company to take at least a few days to pentest newly written code before pushing it live. Yen replied: "That's bullshit and loser thinking. Of course it's possible to be fast and secure." He added: "stop lovingly handcrafting all our code and use AI more."

The investigation drew on eight current employees and one former employee, along with Slack chats and meeting audio. It found staff who held real security concerns but felt unable to raise them because Yen mocks or insults employees in front of colleagues. The cost is now visible. Proton users called for Yen to resign. Yen's response on Reddit was downvoted more than 1,200 times, and many users said they were cancelling paid plans.

Butler conceded the point in the meeting: "Using AI effectively and pragmatically does not mean that we've given up on security, or good architecture, or sound software engineering principles." Later he said: "We fucked up, we know we fucked up."

The engineer who asked for a pentest was doing the work that holds its value. AI does the routine work. You do the thinking. A review that counts how much someone uses AI records the routine half. It has no field for the person who stopped unsafe code going live.

Your Next Move

1. Get the criteria in writing. If your company has floated or dropped an AI assessment, ask your manager this month what the next review will measure: usage, or outcomes. At Proton the measures outlived the event that announced them.
2. Keep a two-column log. In one column, record where AI cut time on real work. In the other, record where you checked its output and caught something. The second column is your evidence of judgement, and it is the column a usage metric misses.
3. Raise risk with a plan attached. When you flag a security or quality problem, bring a scoped check with its time cost already worked out. That gives a leader who is pushing for speed something specific to approve, and it leaves a record that you raised it.