Executive brief
InternLM LMDeploy is a toolkit for deploying and serving large language models (LLMs). The DistServe decode migration feature can be crashed by unauthenticated attackers, causing the inference engine to terminate and rendering the service unavailable to all users. This denial-of-service attack requires no authentication and can be triggered remotely.
Technical details
The vulnerability is a reachable assertion error in the DistServe decode migration loop within the engine_loop.py component. An attacker can send a migration_request with an empty remote_block_ids list, which triggers an AssertionError that crashes the engine loop process. The vulnerability is network-reachable and requires no authentication. Successful exploitation causes an immediate denial of service, as the crashed engine loop prevents all subsequent inference requests from being processed. A patch is likely available in versions after 0.17.0.
Affected products
- InternLM LMDeploy through 0.17.0
Timeline
- 2026-09-17: disclosed