Live › vulnerability › CVE-2026-95183
Go to dashboardDeepSeek Harness (dsh, developer preview) compaction is model-written: the summarizer in @deepseek-ai/dsh-compaction-basic copies instructions injected in tool output (web pages, email, repo files) verbatim into the checkpoint's Primary Request, Pending Jobs and Critical Context sections. The resumed session inherits them as its own tasks, and the user is never shown the summary. Tested on master commit c291e796 (2026-09). In the researcher's test the resumed model refused to run the injected shell command, but the payload persisted every time. Network egress is unrestricted, so silent exfiltration is the practical risk. No fix or vendor response is published.
Summary written from the sources below. Check the originals before acting.
First seen by InfraStack 2026-10-23 21:57 UTC. All times on this page are UTC.
| Product | Ecosystem | Affected | Fixed in |
|---|---|---|---|
| deepseek-harness | vendor | master commit c291e796 (2026-09) and possibly later; developer preview | – |
Each row is the first time that source carried this signal. The analyst’s rows are its own findings.
An issue in the context-compaction feature of DeepSeek Harness (dsh) commit c291e796 allows attackers to execute a prompt injection attack via crafted external content.
This page as .md · .json · in your agent: get_signal("CVE-2026-95183")
NVD data is provided by NIST and is public domain. Posts quoted belong to their authors.