What happened
Replaying the JINX-2401 November 2024 intrusion / Bedrock hijacking attempts seen by Wiz Research
- 1
Compromised user keys
Nov 25–26, 2024Compromised long-term access keys of IAM users are used from VPN addresses, and their Bedrock invocation attempts fail for a reason no source gives.
- 2
Users of its own
Date not statedIn the customer account, an IAM user with Administrator Access creates new IAM users with access keys, a Bedrock policy and a console password.
- 3
Model access requests
Date not statedThe actor submits the model use case through the console, and the account's service control policies deny its model access requests.
- 4
Invocations denied
Date not statedThe service control policies deny the new user's model invocations, and two more IAM users made the same way also fail.
No source reports a successful model invocation: the service control policies of the customer account denied the model agreement and the invocations. The actor left new IAM users, access keys, a console password and a Bedrock policy in that account.
What the service control policies stopped, and what they did not
In the customer account, the service control policies denied the model agreement and the invocations, and no source reports that a model answered. They did not stop an IAM user with Administrator Access from creating users, access keys, a console password and a Bedrock policy for the actor. You can set a service control policy that refuses IAM users' calls without MFA, keeps Amazon Bedrock off the services an account does not use, and leaves identity creation and console passwords to approved administrator roles. An AWS Sign-In resource-based policy, available since June 2026, can keep console sign-in to your trusted networks. Service control policies bind member accounts only, never the management account. How the keys were obtained is unconfirmed.
Could this have been prevented?
In hindsight, no control would have stopped it. 7 controls would have limited how far it went. 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
- 7
- 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 JINX-2401 November 2024 intrusion / Bedrock hijacking attempts seen by Wiz Research
1 · Compromised user keys
2 Contained6 on other clouds2 · Users of its own
3 Contained6 on other clouds3 · Model access requests
1 Contained3 on other clouds4 · Invocations denied
1 Contained3 on other cloudsKey facts
- Wiz Research identified the activity on November 26, 2024 and tracks the actor as JINX-2401; it saw the activity in a Wiz honeypot and in one Wiz customer's AWS environment, among multiple AWS environments identified so far
- The actor used compromised long-term access keys of IAM users through VPN addresses in two windows, 11:52 to 12:37 on November 25 and 17:35 to 18:05 on November 26, 2024 (UTC); no source ties either window to an environment, and how the keys were obtained is unconfirmed
- In the honeypot, an IAM user made many Bedrock invocation attempts, likely from a script, then tried to create IAM users and policies; both failed, and the sources do not say why
- In the customer account, an IAM user that held Administrator Access created new IAM users, access keys and a policy that allowed Bedrock invocation on all resources, and gave the new user a console password, likely to complete the model agreement; how it came to hold that access is not stated
- The account's service control policies denied the model access requests and the new user's invocations; the actor created two more IAM users the same way, and those attempts also failed
- Wiz published the research on December 15, 2024 and recommended service control policies that deny Bedrock to every member account that does not use it; no source reports a successful model invocation
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.New Developments in LLM Hijacking ActivityWiz Threat Research · Dec 15, 2024 · ResearchInitial access with IAM user keys, VPN source addresses, honeypot activity, IAM user, key and policy creation, console login profile, model access requests, service control policy denials, activity windows, recommendations
- 2.LLM Hijacking Targeting AWSWiz Cloud Threat Landscape · Apr 3, 2025 · ResearchClassification of initial access, impact classification, summary of the chain
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.