Incident analysis

One laptop, $6.3 billion: what a compromised engineer machine can reach

Binance says its breach may have started on a DevOps engineer's laptop. We do not know what that laptop held. We do know what most engineers' laptops hold, and how much of the year it can reach production.

A compromised laptop can do whatever the access on it can do, whenever that access is live. Device security lowers the odds of the compromise. Time-bound access shrinks what follows: on an ordinary day the laptop's account belongs to no production group, and getting into one means asking somebody else.

What Binance has said

On 29 October 2026, about $6.3 billion in customer assets left Binance's wallets. By most counts it is the largest crypto theft on record. Binance's initial investigation said the incident may have started with a compromised DevOps engineer's laptop, which gave an attacker access to a service container, and unauthorized withdrawal transactions were then submitted from that container. Binance said it found no evidence that a wallet private key leaked.

Within hours that became a more careful story. Co-CEO Richard Teng said the attacker did not simply use one engineer's permissions and that it was not a one-click action, and the exchange said investigators had found further points of breach in what it called a complex attack. About $3 billion has since been frozen, and Binance has said it will cover customer losses. Outside firms have linked the activity to a North Korean group. Binance has not attributed it.

What we are not claiming

That TemprBac would have stopped this. Nobody outside Binance knows what that laptop could reach, and the withdrawals came from a service container, not a person. This post is about the pattern the first finding points at, which is common and worth fixing whoever is next.

The pattern: access that is always on

Most engineering teams grant production access once and then leave it in place. A cloud role in a group the engineer joined when they started, a kubeconfig that never expires, an SSH key, a browser session good for weeks. None of that is negligence. It is the fastest way to let people do their jobs, and it holds until the laptop it lives on belongs to somebody else.

Then three things are true at once:

  • The window is always open. If the engineer can reach production at any hour, so can whoever controls their machine, including at 3 a.m. when nobody is watching.
  • The activity looks legitimate. Everything runs as a real engineer with real permissions, so the logs show someone working.
  • The access outlives the need. A permission added for one incident or one migration is rarely taken away, so what a stolen laptop reaches grows year after year.

Phishing-resistant MFA, endpoint detection and device management all make a compromised laptop less likely. None of them change what an attacker can do once they have one.

The same week, held two ways

A grid of the 168 hours in a week, one square per hour. With standing access every square is filled, so the laptop can reach production all 168 hours, including Wednesday at 3 a.m. With two four-hour grants, on Tuesday morning and Thursday afternoon, 8 squares are filled and the other 160 hours reach nothing.
An engineer who needs production twice a week. With standing access the laptop can reach it every hour. With time-bound grants it reaches it for eight hours, and Wednesday 3 a.m. is not one of them.
8,760 hours a year a standing role is live 24 × 365. Nights, weekends and holidays included.
416 hours a year with two four-hour grants a week An example, not a promise: 8 hours a week for 52 weeks, under 5% of the year.
0 hours on an ordinary day With no grant open, the account belongs to no production group at all.

Your own number depends on how often your engineers really need production, and finding it out is worth doing anyway. Look at the last month of production changes, count the people who made one, and compare that with the people who could have.

How a grant works in TemprBac

TemprBac does not replace your identity provider or your cloud. It manages group membership in the systems you already run (Entra ID, Okta, Google Workspace, GitHub teams, AWS IAM groups and Active Directory), so the groups that carry production rights have members only while a grant is open.

  1. Nothing standing. The production groups are empty by default, and that includes TemprBac's own admin rights, which are a time-bound package like any other. A laptop on an ordinary day reaches nothing through them.
  2. Request a package, with a reason. A package is a named bundle of group memberships, such as "production database, read and write". The engineer asks for it, says why, and the request goes to that package's named approvers by email, and to a Slack channel or Teams workflow if the package is set to notify one. An unanswered request closes itself after 24 hours.
  3. A window fixed before it opens. Each package has its own window, from 1 to 8 hours (4 by default). Removal is scheduled at the moment of the grant, so nobody has to remember to do it and nothing waits for the engineer to log off.
  4. A record of each step. Who asked, what for, who approved, when the access arrived and when it was removed, all in one place. That makes the request nobody expected easy to spot.

What an attacker on the laptop has to do now

An intruder who controls an engineer's machine also controls their TemprBac sign-in. So to reach production, they have to file a request, and the request is the loudest thing they can do:

From the attacker's sideStanding roleTime-bound grant
When the laptop reaches production Every hour of the year Only while a grant is open
To reach it at 3 a.m. Use what is already there Ask a named approver, at 3 a.m., in the name of an engineer who is asleep
What the record shows An engineer doing normal work A request, its stated reason, a decision, a grant and a removal, each with a time
When the access ends When somebody notices When the window closes, 1 to 8 hours after the grant

This works only if the packages that matter need a second person. An auto-approved package grants to whoever asks, attacker included, so keep auto-approval for read-only access, where the window and the record are the control. If your org turns on break-glass, a requester can approve their own request after an unanswered wait of 5 to 60 minutes. Their approvers and admins are told at once, which is the point, and we have written up what that trade costs.

What just-in-time access does not solve

A shorter window is still a window. Here is where it ends:

  • During an open grant, the attacker has what the grant has. If they are on the laptop while the engineer works, they work too. Narrow packages limit how much that is.
  • Workloads are not people. Binance's withdrawals were submitted from a service container. A container or a CI job that holds its own long-lived credential is outside any human approval, and needs short-lived workload identity of its own. TemprBac grants people group membership. It does not issue workload credentials.
  • Sessions issued during a grant can outlive it. Removing a membership stops new access. A cloud session or token already handed out may stay valid until it expires, so keep those lifetimes short as well.
  • An approver's laptop is worth more than a requester's. An attacker who holds both can approve their own request. Device security for approvers matters more, not less.
  • The last control is on the action. Withdrawal limits, multi-party signing and rate limits on high-value operations decide how much a single compromised path can move, whoever is holding it.

Just-in-time access belongs beside those controls, not instead of them. Its job is to make sure that most of the time, a compromised laptop is only a laptop.

Questions

Was the Binance hack caused by an engineer's laptop?

Binance's initial investigation named a compromised DevOps engineer's laptop as the possible entry point. It has since said the attack had more than one point of breach and did not rest on one engineer's permissions. The full root cause is not public yet.

How much was taken?

About $6.3 billion, according to Binance, and roughly $3 billion of it has been frozen. Binance has said it will cover customer losses.

What is just-in-time access?

Access that is granted when someone needs it, for a fixed window, and removed automatically when the window ends, instead of permissions that stay on an account indefinitely. Our guide to just-in-time access covers how it works and why the removal is the hard half.

Would it stop every laptop-based attack?

No. It limits when a compromised device can reach production, and makes reaching it visible. It works best with device security, phishing-resistant MFA, short-lived workload credentials and controls on high-risk actions.

How is this different from ordinary RBAC?

Role-based access control decides what a role can do. Time-bound access also decides when: the role is held for a short, approved window and then it is gone.

Sources

How many hours a year can your laptops reach production?

TemprBac makes that number the hours your engineers are actually working in production, across the systems you already run, with a record of every one of them.

Join the waitlist