These pages are published after PatchSiren validates generated defensive summaries against stored public CVE and source evidence.
CVE-2026-105135 is a critical vulnerability in InternLM MindSearch 0.1.0, affecting the Planner Agent component. The issue allows for remote code injection via the ExecutionAction.run function in mindsearch/agent/graph.py. The CVE record was published on 2026-10-04T07:16:33.693Z and has not been modified since then. The NVD entry is currently Received.
CVE-2026-92983 debrief: InternLM LMDeploy vulnerability allows unauthenticated attackers to cause memory exhaustion via session ID mismatch. The vulnerability exists in the DistServe prefill/decode disaggregation mode, where the proxy uses user-facing session IDs instead of internal scheduler keys, leading to unreleased scheduler sessions and potential memory exhaustion. Defenders should assess exposure a [truncated]
CVE-2026-92971 InternLM LMDeploy through 0.17.0 contains a reachable assertion vulnerability in the DistServe decode migration loop that allows unauthenticated attackers to terminate the inference engine. This vulnerability can be triggered by submitting a migration_request with an empty remote_block_ids list, causing an AssertionError that crashes the engine loop and results in subsequent inference reque [truncated]
CVE-2025-59953 debrief: LMDeploy rpc server deserialization vulnerability allows remote code execution. Affected product deployments should assess exposure and apply the patch in version 0.10.2. The vulnerability exists in LMDeploy's rpc server (AsyncRPCServer in zmq_rpc.py) due to insecure deserialization practices. Defenders should verify exposure, apply the patch, and monitor for potential exploitation [truncated]
CVE-2026-76850 is a critical vulnerability in LMDeploy, a project by InternLM, that allows for remote code execution via unsafe pickle deserialization in the disaggregated serving peer connector. The vulnerability exists in the handle_zmq_recv coroutine in lmdeploy/pytorch/disagg/conn/engine_conn.py, where peer-to-peer cache-free requests are deserialized using pickle.loads() without proper authentication [truncated]
The lmdeploy OpenAI-compatible API server has a server-side request forgery (SSRF) vulnerability. This vulnerability allows unauthenticated attackers to access internal services and cloud metadata endpoints by supplying a crafted image_url that redirects to internal targets. The vulnerability exists because the API server follows HTTP redirects without re-validating each hop through the safety guard. This [truncated]
CVE-2026-46517 is a high-severity vulnerability in LMDeploy, a toolkit for compressing, deploying, and serving large language models. The vulnerability has a CVSS score of 7.8 and is classified as HIGH. It was published on 2026-06-10T00:16:53.827Z and modified on 2026-06-11T12:16:31.507Z. The vulnerability is caused by hardcoded 'trust_remote_code=True' which enables HF supply-chain RCE without user opt-i [truncated]
LMDeploy, a toolkit for compressing, deploying, and serving large language models, is vulnerable to arbitrary code execution. This vulnerability is due to hardcoded 'trust_remote_code=True' in multiple HuggingFace model-loading call sites in versions 0.12.3 and prior. The vulnerability has a CVSS score of 7.8 and is classified as HIGH severity.