What happened
Replaying the May 2026 marimo intrusion by a suspected LLM agent / observed by Sysdig on May 10, 2026
- 1
Notebook shell and credential harvest
May 10, 2026, 18:23–18:24 UTCAn unauthenticated terminal flaw gives a shell on the notebook host, and two cloud credentials are read from its files and environment.
- 2
Cloud key and secret
May 10, 2026, 19:26 UTCThe first harvested key passes an identity check from Cloudflare Workers addresses, and Secrets Manager returns an SSH private key.
- 3
Bastion and database
May 10, 2026, 19:30–19:32 UTCThe SSH key opens a bastion, and a password read there opens an internal database whose six tables are dumped.
An SSH private key was read from AWS Secrets Manager, a bastion was entered and six tables of an internal PostgreSQL database were dumped, in under one hour from the exploit.
Where the provider's controls fit
On AWS, a security group, or a declarative policy that sets VPC Block Public Access to ingress-only, bounds who can reach your notebook from the internet, and a security group bounds who can reach SSH on your bastion, where they run in an AWS VPC; the source does not say where they ran, so it is not known whether these connections would have been refused. A resource control policy that keeps Secrets Manager to your own networks bounds the reads of the secret to those networks in a member account; the source does not name the account. A service control policy that refuses calls from outside your networks, or IAM user calls without MFA, would have bounded the key, but not its identity check, which needs no permission. These organization policies need AWS Organizations. No cloud control reaches the missing login check in marimo, the files read on the host, or a database login made with its own password from the bastion.
Could this have been prevented?
In hindsight, no control would have stopped it. 8 controls would have limited how far it went. 3 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.
- 0
- Prevented
- 8
- Contained
- 3
- 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.
Replaying the May 2026 marimo intrusion by a suspected LLM agent / observed by Sysdig on May 10, 2026
1 · Notebook shell and credential harvest
3 Contained2 Not covered6 on other clouds2 · Cloud key and secret
4 Contained6 on other clouds3 · Bastion and database
1 Contained1 Not covered3 on other cloudsKey facts
- On May 10, 2026 at 18:23:44 UTC, a connection to the terminal endpoint of an internet-reachable marimo notebook gave a shell without authentication (CVE-2026-39987); marimo had published the advisory and the 0.23.0 fix on April 8, 2026
- From 18:24:14 UTC, application environment files, the process environment and the AWS credentials file on the host were read, and two cloud credentials were taken; whether they were long-term access keys is not stated
- From 19:26:31 UTC the first key made an identity check, listed the Secrets Manager secrets and read an SSH private key, in 12 calls over 22 seconds spread across 11 Cloudflare Workers locations; the source is inconsistent about the gap before the first AWS call
- Eight SSH sessions from six Cloudflare Workers addresses opened a bastion with that key between 19:30:30 and 19:32:23 UTC
- A database password read from a file on the bastion opened an internal PostgreSQL database, and six tables were dumped at 19:32:23 UTC, under one hour after the exploit
- Sysdig attributed the operation to an LLM agent from four indicators that it does not present as proof, and it did not say who owned the environment or whether it was a sensor
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.AI agent at the wheel: How an attacker used LLMs to move from a CVE to an internal database in 4 pivotsSysdig · May 26, 2026 · ResearchEntry vector, credential harvest, AWS calls, egress pattern, bastion sessions, database dump, timeline
- 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.