Summary
JSON objects after decoding might use more memory than their serialized version. It is possible to tune a JSON to maximize the factor between serialized memory usage and deserialized memory usage (similar to a zip bomb). While reproducing the issue, we could reach a factor of about 35. This can be used to circumvent the [max_request_size (https://openbao.org/docs/configuration/listener/tcp/) configuration parameter, which is meant to protect against Denial of Service attacks, and also makes Denial of Service attacks easier in general, as the attacker needs much less resources.
Details
The request body is parsed into a map[string]interface{} https://github.com/openbao/openbao/blob/788536bd3e10818a7b4fb00aac6affc23388e5a9/http/logical.go#L50 very early in the request handling chain (before authentication), which means an attacker can send a specifically crafted JSON object and cause an OOM crash. Additionally, for simpler requests with large numbers of strings, the audit subsystem can consume large quantities of CPU.
To remediate, set max_request_json_memory and max_request_json_strings.
Impact
- Unauthenticated Denial of Service
 
Resources
This issue was disclosed directly to HashiCorp and is the OpenBao equivalent of the following tickets:
HashiCorp attributes the problem to the audit subsystem. For OpenBao, it was noted the problem was additionally in the requests handling logic.
References
   
 
Summary
JSON objects after decoding might use more memory than their serialized version. It is possible to tune a JSON to maximize the factor between serialized memory usage and deserialized memory usage (similar to a zip bomb). While reproducing the issue, we could reach a factor of about 35. This can be used to circumvent the [
max_request_size(https://openbao.org/docs/configuration/listener/tcp/) configuration parameter, which is meant to protect against Denial of Service attacks, and also makes Denial of Service attacks easier in general, as the attacker needs much less resources.Details
The request body is parsed into a
map[string]interface{}https://github.com/openbao/openbao/blob/788536bd3e10818a7b4fb00aac6affc23388e5a9/http/logical.go#L50 very early in the request handling chain (before authentication), which means an attacker can send a specifically crafted JSON object and cause an OOM crash. Additionally, for simpler requests with large numbers of strings, the audit subsystem can consume large quantities of CPU.To remediate, set
max_request_json_memoryandmax_request_json_strings.Impact
Resources
This issue was disclosed directly to HashiCorp and is the OpenBao equivalent of the following tickets:
HashiCorp attributes the problem to the audit subsystem. For OpenBao, it was noted the problem was additionally in the requests handling logic.
References