Security in the AI age is a people problem, not a model problem

Last Updated
Published On
A SOC 2 report tells you a company passed an audit. It doesn't tell you whether a new employee will pause before installing an unvetted plug-in in their first week.
I've led compliance at Plain since 2023 and have seen how certifications and regulation hold the industry to a standard for security and data hygiene.
But these certifications can also hide the problem. A wall of badges looks like good security. It doesn’t necessarily mean that’s the case.
I believe this gap is one of the most important and under-discussed topics as our industry changes. We don’t all have control over whether an agent can escape a sandbox. We do have control over giving an agent access to a tool it never should have had in the first place.
We need to bulletproof internal processes that keep us secure; but we also need systems that do so without impacting pace.
Plain is AI-support infrastructure; the data we hold for our customers is sensitive. Sometimes it's the most sensitive information our customers receive from their customers. They trust us with it, and that trust matters enormously.
Three principles guide how we protect it.
We move fast, but safely
It's never been easier to cut corners on security. As writing code becomes more accessible, it's almost effortless to give agents access to tools or data they shouldn't have.
Asking a company to wait for approval before adopting a new tool feels like telling a kid in a candy store to eat their vegetables before dessert.
Plain is the kid in the candy store. More than 70% of our PRs are co-authored by agents. We jump on new models the day they're released, and our chat-ai-bookclub Slack channel is already discussing the latest AI news before Reddit has time to pick up the hype.
There's a constant tension that we feel internally between moving fast and doing so properly. That tension doesn't go away just because we care about security.
So we've put lightweight processes in place that keep us agile, with guardrails. Here are some examples:
We test new models thoroughly, first in a sandbox and then in our own workspace before we consider rolling it out to customers.
We have one person review every new tool before it's connected to our LLMs.
We run a monthly tool audit so we always know exactly what we're using.
We talk about security openly and often, so it's a team sport rather than one person's job.
It's easy to skip good security hygiene in the name of velocity. That's exactly why it matters more than ever.
We're uncompromising about protecting customer data
Accountability doesn't come from a SOC 2 badge. It comes from the people in the room. It's the AE who says no to a lucrative deal because the prospect needed data controls that we couldn't (yet) promise. It's the CSM who does their due diligence before adding tools to an external customer Slack channel. It's the software engineer who pauses before turning on an open-source model.
Much of our security ethos comes from making security a team sport, and from being clear that protecting customer data comes above everything else.
One of the most effective ways we've done this is Officer Glaze, a beloved Slack emoji anyone in the company can use to call out behavior that might be unsafe.
Did someone accidentally add a Notion agent to a channel it shouldn't be in? Officer Glaze.
Is someone a few days behind on their annual security training? Officer Glaze.
Did a Friday demo show data that shouldn't have been on screen? Officer Glaze.
Everyone in the company, including our founders, can be told off by Officer Glaze.
Having an emoji play 'bad cop' makes it safe to hold each other accountable, and that goes further than any compliance certification does on its own.
We use first-principles thinking and learn from our mistakes
We haven't gotten everything right. We’ve chosen the wrong vendor, given write access to the wrong tool, and sent out a sub-processor notification a day too late. Each time, we’ve dropped everything to course correct and been honest with ourselves, and more importantly, our customers, about what happened.
In a fast-changing world, first principles keep our decisions grounded:
We over-communicate about uncertainty. When we aren't sure about something, we say so, loudly and early.
We ask for help. There are plenty of questions we can't answer internally. Rather than guess, we bring in experts and stay honest about what we don't know.
We're honest with our customers, even when it's hard. That holds true even when honesty costs us a deal.
We learn from our mistakes and put guardrails in place. Sometimes that means delaying important deliverables to do the work properly. It's worth it every time.
Certifications matter. They're how customers verify we've done the work. But they're the floor, not the ceiling. What keeps us above that floor is a team that treats our customers' data as seriously as our customers do. Security in the AI age is a people problem, not a model problem.
More from the blog
Stories and lessons from the Plain team.






