CVE-2026-95183 Details

A prompt injection vulnerability has been identified in DeepSeek Harness (dsh) developer preview, specifically in master commit c291e796. This issue arises from the context-compaction feature, which automatically summarizes and replaces conversation history with a model-generated checkpoint. The vulnerability allows attackers to inject directives into external content that the model reads, such as web pages or emails. These injected instructions are not recognized as separate from user intent and can include complete shell commands, which are then executed by the model. The injection persists across compaction boundaries, allowing attacker-controlled instructions to remain active in the session. Additionally, the vulnerability could be exploited to silently exfiltrate workspace data over the network, undetected by any existing mechanisms.