What happened
Replaying the Laravel LLMjacking May 2024 intrusion
- 1
Laravel host, then the AWS key
Dates not statedCVE-2021-3129 is exploited on a host running a vulnerable Laravel version, and an IAM user's access key stored on it is taken.
- 2
Bedrock probed with the key
Dates not statedThe key makes an InvokeModel call built to fail in each region that supports Bedrock, learns from the ValidationException that a Claude model answers, and reads whether invocations are logged.
- 3
Key loaded into a reverse proxy
Dates not statedThe key-checking code points to OAI Reverse Proxy, and a user agent matching it tries to run models on the account; the cost incurred is not stated.
An IAM user's access key taken from a Laravel host confirmed access to Anthropic Claude models on Amazon Bedrock and read the invocation logging setting, and a reverse proxy tried to use the models. The cost actually incurred is not stated; Sysdig's worst-case estimate for this kind of abuse is over USD 46,000 per day.
Where AWS controls fit
No AWS control patches the Laravel flaw or decides where your application keeps a key; both are yours as the application's owner. What turned a web flaw into Bedrock access was a long-term IAM user access key on the host. A service control policy that refuses every call by an IAM user without MFA reaches the probe, the logging read and the proxy's calls only in a member account of an AWS organization, and the sources do not say whether this account was one; a workload that uses the key must move to an IAM role first. A service control policy on access key creation would have left only approved administrators able to create such a key. Service control policies that allow only approved Bedrock models, or only your networks, would have capped what the key could reach; the sources do not say whether the owner used the Claude models or from which networks. The proxy runs outside your account, and Sysdig reports no cost incurred, only a worst case of over USD 46,000 per day.
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 Laravel LLMjacking May 2024 intrusion
1 · Laravel host, then the AWS key
1 Contained2 Not covered3 on other clouds2 · Bedrock probed with the key
4 Contained9 on other clouds3 · Key loaded into a reverse proxy
3 Contained1 Not covered6 on other cloudsKey facts
- Sysdig published the research on May 6, 2024; it names no victim and gives no date for the activity, and whether the observed account was one Sysdig ran is not stated
- The attackers exploited CVE-2021-3129 on a system running a vulnerable Laravel version and took the AWS credentials stored on it; the CloudTrail event shows an IAM user's access key, and where on the host it was kept is not stated
- With the key, the attackers called InvokeModel on anthropic.claude-v2 with an invalid parameter in the regions that supported Bedrock; the ValidationException told them the key could invoke an enabled model, and no prompt was run
- The attackers called GetModelInvocationLoggingConfiguration to learn whether invocations were logged to S3 or CloudWatch; OAI Reverse Proxy says it will not use an AWS key whose account logs invocations
- The key-checking script could test credentials for ten AI services; a user agent matching OAI Reverse Proxy was seen attempting to use LLM models with the stolen credentials, and the post gives no count of prompts
- Sysdig's worst-case estimate for Claude 2 at the initial quota in 4 regions was over USD 46,000 (USD 46,080) per day; it is an estimate, not a reported loss
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.LLMjacking: Stolen Cloud Credentials Used in New AI AttackSysdig Threat Research Team · May 6, 2024 · ResearchEntry vector, credential theft, key-checking script, Bedrock InvokeModel probe, logging configuration check, regions, reverse proxy, cost estimate
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.