Opal Security

It is 2 a.m., and a piece of software just asked for the keys to your database.

The request looks routine. An AI agent, spun up to triage support tickets, needs read access to a customer table. No human is awake to approve it. A decade ago this scenario did not exist. Today it happens thousands of times a night inside the average enterprise, and somebody - or something - has to answer. At Opal Security, that answer is automatic, scoped, time-boxed, and logged. The agent gets exactly what it needs, for exactly as long as it needs it, and not a permission more.

That is Opal in 2026: a San Francisco company of 62 people that has quietly appointed itself the bouncer for every identity in the modern org chart - employees, service accounts, and now the fast-multiplying population of AI agents. It is not the flashiest job in security. It is arguably the most important one nobody wanted.

Access has a way of accumulating like junk in a garage.

Here is the uncomfortable truth most companies prefer not to audit: people collect access and never give it back. The intern who needed production logs for one afternoon in 2021 still has them. The engineer who switched teams kept the old permissions and gained new ones. Multiply that by a few thousand employees, a few hundred apps, and a cloud account nobody fully maps, and you get what the industry politely calls "access sprawl." It is the security equivalent of leaving every door in the building unlocked because checking them all is exhausting.

The traditional fix was the access review - a quarterly ritual in which managers squint at spreadsheets and approve everything because saying no requires understanding, and understanding requires time nobody has. It is governance theater. The boxes get checked. The doors stay open.

Translation: most breaches don't kick the door down. They walk in with a badge somebody forgot to deactivate.

An MIT cryptographer and a Dropbox engineer walk into the least glamorous problem in tech.

Umaimah Khan studied mathematics and computer science at MIT, did defense research embedded with government agencies including DARPA, then worked at Collective Health and Amplitude. Stephen Cobbe came from engineering at Dropbox. In 2020 they co-founded Opal around a deeply unsexy observation: nobody actually knows who has access to what, and the tools meant to answer that question were built for a world before cloud, before microservices, and certainly before AI agents.

Their bet was contrarian. Instead of building yet another dashboard to look at access - the industry already had plenty of those - they would build infrastructure to govern it as code. Define who should have what in policy. Enforce it continuously. Grant access just in time, then claw it back automatically. Treat least privilege not as an aspiration printed in a compliance doc, but as a default the system maintains on its own.

It was the kind of bet that sounds obvious in retrospect and reckless at the time. Selling "we'll take away access you didn't know you had" is not a pitch that lights up a room. But the people who run security for a living understood immediately, because they were the ones losing sleep.

Five tools, one stubborn idea: access should expire by default.

Opal's platform is built around the belief that standing privilege is the enemy. Most access shouldn't be permanent; it should be requested, granted for a window, and revoked when the window closes. Around that spine, the company assembles a handful of capabilities that work together rather than as bolt-ons.

What you can actually do with it: hand an engineer database access for two hours, let it self-destruct, and have a clean audit trail with zero Slack threads.

Secure every identity and access path. Even the ones that aren't people.

Opal's mission statement is mercifully free of jargon: secure every identity and access path. What has changed is the meaning of the word "identity." When the company started, identity meant employees and the occasional service account. By 2026, it means a workforce of software - AI agents that request, act, and need governing exactly like the humans beside them, only faster and at far greater volume.

This is where Opal's early bet pays a second dividend. A platform built to grant and revoke access programmatically does not care whether the requester has a pulse. The same just-in-time logic, the same policy-as-code, the same audit trail - it all applies to an agent. In December 2025 the company brought in Howard Ting as CEO, and in mid-2026 raised $23M alongside five senior hires, explicitly to unify governance across human, non-human, and agentic identities. The org chart, in other words, is being rewired for a workforce that doesn't sleep.

The number of things asking for access is about to explode.

Every company racing to deploy AI agents is, whether it realizes it or not, also racing to create thousands of new identities with credentials, permissions, and the ability to act. The old model - quarterly reviews, standing access, trust by default - was already failing for humans. Against a machine workforce that scales in minutes, it doesn't stand a chance.

That is the wager underneath Opal's whole story. Governance can't be a meeting anymore. It has to be infrastructure - code that grants, watches, and revokes on its own, indifferent to whether the requester is a person or a process. Opal has spent six years building exactly that, which is either very good timing or very good instinct. Possibly both.

So return to that 2 a.m. request. The AI agent asks for the customer table. A few years ago, the honest answer would have been a shrug - grant it, hope for the best, find out later. Now the request is met by a policy that already knew this might happen, a grant that expires before sunrise, and a record that will still make sense at the next audit. Nobody woke up. Nothing leaked. The door opened just wide enough, and then it closed.

The keys are still in the building. Opal just made sure they stop walking out.