Skip to main content
Back
IntrusionAWS
9 min read

Hand-operated marimo notebook exploit to AWS Secrets Manager and a bastion host

Victim
An unnamed AWS account behind an internet-reachable marimo container watched by Sysdig
When
not stated; between April 8 and September 11, 2026Disclosed Sep 11, 2026
Entry
Unauthenticated remote code execution through the terminal WebSocket of an internet-reachable marimo container (CVE-2026-39987).

The Sysdig Threat Research Team watched an intrusion that began at the unauthenticated terminal endpoint of an internet-reachable marimo container (CVE-2026-39987) and gives its session times in UTC with no calendar date. Two AWS access keys of two different IAM users were taken, one from the host's environment and credential files and one from the application's Redis backend; a script on the host read an SSH private key from AWS Secrets Manager, and that key authenticated to an internet-reachable bastion twice, while EC2 enumeration and an EC2 Instance Connect key push were refused. A TCP sweep of the host's private network and a reverse-shell listener left no stated result, Sysdig reports no data exfiltration, and it does not say who owned the AWS account or where the hosts ran.

What happened

External Parties
Disposable VPS
Attacker-owned VPS listener
Sysdig Threat Research
Internet
AWS APItwo IAM user keys
marimo
Internet-reachable marimo host
AWS Account
+ SSH bastion
+ key from Secrets Manager
+ no command described
Internet-reachable bastion
marimo Container
Redis Backend
Credential Files
Secrets Manager Secret
EC2 API

Replaying the 2026 hand-operated marimo intrusion / session times in UTC; no calendar date published

  1. 1

    Notebook shell

    Date not stated, 12:52–12:54 UTC

    An unauthenticated terminal WebSocket gives a shell on the marimo container, and the first command sweeps the host's private network.

  2. 2

    Two AWS keys harvested

    Date not stated; before the first session, and 18:57 UTC

    An AWS access key is read from the host's environment and credential files, and the application's Redis backend returns a second one.

  3. 3

    AWS keys, secret and EC2

    Date not stated; keys checked earlier, calls 16:50–18:56 UTC

    Both keys pass an identity check, a script reads an SSH private key from Secrets Manager, and EC2 enumeration and a key push are refused.

  4. 4

    Bastion and reverse shell

    Date not stated, 16:51–21:50 UTC

    The SSH key from Secrets Manager opens an internet-reachable bastion twice, and a reverse-shell listener is set up with no confirmed connect-back.

Impact

An operator took two AWS access keys of two IAM users from an internet-reachable marimo container and its Redis backend, read an SSH private key from AWS Secrets Manager and authenticated to an internet-reachable bastion. EC2 enumeration and an EC2 Instance Connect key push were refused. Sysdig reports no data exfiltration and describes no command on the bastion.

Outcome

Where the provider's controls fit

On AWS, the combined inbound rules of all security groups on your notebook host admitting the notebook port only from your own address ranges and excluding the attack source, or a declarative policy that sets VPC Block Public Access to ingress-only across your organization, would have refused the connection from the internet to the notebook's terminal, if the host runs in an AWS VPC that no exclusion covers; the source does not say where it ran. The combined rules of all security groups on your bastion admitting SSH only from your own ranges would have refused the connection that used the stolen key, if the bastion runs in an AWS VPC and none of those groups allows the notebook host. A security group with no outbound rule to the internet bounds where your notebook host can connect out, where it runs in an AWS VPC. A service control policy that refuses IAM user calls without MFA, a resource policy on your secret and a resource control policy that keeps Secrets Manager to your own networks each bound what a stolen IAM user key can read, in a member account of an organization; the source does not say the account was one. No policy refuses the key's identity check, which needs no permission. No cloud control reaches the missing login check in marimo, the files and the Redis store read on your host, or the bastion's own SSH server.

Could this have been prevented?

In hindsight, not as a whole. 3 controls would have refused a step outright, and 7 controls would have limited one. 5 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

3
Prevented
7
Contained
5
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

External Parties
Disposable VPS
Attacker-owned VPS listener
Sysdig Threat Research
Internet
AWS APItwo IAM user keys
marimo
Internet-reachable marimo host
AWS Account
+ SSH bastion
+ key from Secrets Manager
+ no command described
Internet-reachable bastion
marimo Container
Redis Backend
Credential Files
Secrets Manager Secret
EC2 API

Replaying the 2026 hand-operated marimo intrusion / session times in UTC; no calendar date published

1 · Notebook shell

2 Prevented1 Contained1 Not covered6 on other clouds

2 · Two AWS keys harvested

1 Contained2 Not covered3 on other clouds

3 · AWS keys, secret and EC2

4 Contained1 Not covered9 on other clouds

4 · Bastion and reverse shell

1 Prevented1 Contained1 Not covered6 on other clouds

Key facts

  • marimo published the advisory for the unauthenticated terminal WebSocket (CVE-2026-39987) and the 0.23.0 fix on April 8, 2026; the intrusion's sessions carry no calendar date in Sysdig's write-up of September 11, 2026
  • At 12:52:18 UTC a WebSocket connection from one source address that Sysdig calls a likely disposable VPS gave a shell on the marimo container, and at 12:54:13 UTC the first command swept the host's private /24 network; what the sweep found is not stated
  • Two AWS access key pairs, mapped to two different IAM users, were taken from the host's environment and credential files and from the application's Redis backend, and both were validated with STS GetCallerIdentity, the first more than 28 hours before the first observed session
  • At 16:50:40 UTC a boto3 script on the host called Secrets Manager GetSecretValue and got an SSH private key; which of the two keys made the call is not stated
  • Three EC2 calls were denied and one EC2 Instance Connect SendSSHPublicKey call was blocked from 18:56:32 to 18:56:50 UTC, while the SSH key authenticated to an internet-reachable bastion at 16:51:45 and 18:57:30 UTC; no command run on the bastion is described
  • Sysdig reported more than 850 interactive commands over about nine hours, a reverse-shell listener with no confirmed connect-back and no data exfiltration, and it assessed a single human operator from behavior, not proof

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    Machine speed, hold the AI: Hand-rolled marimo CVE-2026-39987 exploitSysdig · Sep 11, 2026 · ResearchEntry vector, credential sources, AWS calls and their outcomes, bastion access, operator profile, timeline
  2. 2.
    Pre-Auth Remote Code Execution via Terminal WebSocket Authentication Bypassmarimo · Apr 8, 2026 · AdvisoryVulnerability, affected versions, fix

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns