What happened
Replaying the AI-assisted AWS intrusion, November 2025
- 1
Exposed credentials
Nov 28, 2025, offset 0:00The credentials of an IAM user are read from public S3 buckets that hold RAG data for AI models; how the buckets became public is not stated.
- 2
Key use and role assumption
Nov 28, 2025, offsets 0:00–0:06The credentials are used from VPN and rotating addresses to call about 10 services, and roles named sysadmin, netadmin and account are assumed; why the assumptions succeeded is not described.
- 3
Admin through Lambda
Nov 28, 2025, offsets 0:08–0:11 and 1:00The code of an existing Lambda function is replaced three times; the injected code creates an access key for an admin user, a backdoor admin user follows at offset 0:11, and keys for two more users follow at 1:00.
- 4
Model and compute abuse
Nov 28, 2025, offsets 0:58–1:51Bedrock models are invoked 13 times with invocation logging disabled, secrets, parameters and logs are read, and a GPU instance runs for about 5 minutes.
Over 1 hour 51 minutes the actor assumed six IAM roles, took over four IAM users with new access keys and created one, read Secrets Manager secrets, SSM parameters, CloudWatch logs and internal S3 data, invoked Amazon Bedrock models 13 times and launched a GPU instance that ended after about 5 minutes; the cost to the victim and the actor's goal are not stated.
Where cloud-provider controls fit
On AWS, S3 Block Public Access would have kept the credentials out of the internet's reach, and a resource control policy bounds the buckets to principals of the organization. With the key in hand, a service control policy that limits the caller's network and a permissions boundary on the automation user would have capped what the key reaches, and a service control policy on lambda:UpdateFunctionCode would have refused the code change. A Lambda code signing configuration set to Enforce bounds what can be deployed for as long as it stays on the function. A service control policy that lists the approved Bedrock models bounds what an admin identity can still call. The sources do not say why the first role assumptions succeeded, so the role trust policy row rests on policies they do not describe.
Could this have been prevented?
In hindsight, not as a whole. 4 controls would have refused a step outright, and 12 controls would have limited one. 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.
- 4
- Prevented
- 12
- Contained
- 0
- 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 AI-assisted AWS intrusion, November 2025
1 · Exposed credentials
1 Prevented1 Contained3 on other clouds2 · Key use and role assumption
1 Prevented3 Contained6 on other clouds3 · Admin through Lambda
1 Prevented4 Contained9 on other clouds4 · Model and compute abuse
1 Prevented4 Contained9 on other cloudsKey facts
- Sysdig observed the intrusion on November 28, 2025 and published on February 3, 2026; the post gives no time of day, so its timeline runs in offsets from 0:00 and ends at 1:51
- The credentials of an IAM user with ReadOnlyAccess and read and write permissions on Lambda were found in public S3 buckets that held RAG data for AI models; how the buckets became public is not stated
- At offset 0:08 the code of an existing Lambda function was replaced in three iterations and its timeout raised from 3 to 30 seconds; the injected code created an access key for an admin user and returned it in the function output, and at 0:11 a new user was created with AdministratorAccess attached
- Six IAM roles were assumed across 14 sessions and five IAM users were involved, four taken over with new access keys and one created; Sysdig counts 19 unique principals and the post does not show how it derives the figure
- Between offsets 0:58 and 1:51 Bedrock models were invoked 13 times after model invocation logging was found disabled; at 1:42 a p4d.24xlarge instance launched at about $32.77 an hour and ended after about 5 minutes, for a reason the researchers did not determine
- AWS said, as The Register reports, that its services and infrastructure were not affected and that the account was compromised through misconfigured S3 buckets
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-assisted cloud intrusion achieves admin access in 8 minutesSysdig Threat Research Team · Feb 3, 2026 · ResearchTimeline, initial access, privilege escalation, persistence, Bedrock abuse, GPU launch, principals, evidence of AI use, recommendations
- 2.AWS intruder achieved admin access in under 10 minutes thanks to AI assist, researchers sayThe Register · Feb 4, 2026 · NewsSummary of Sysdig findings, 8 versus 10 minutes, AWS statement
- 3.AI-Assisted Cloud Intrusion: Admin Access in 8 MinutesCloud Security Alliance · Mar 15, 2026 · ResearchAttribution caution, recommendations
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.