CVE-2026-95182: DeepSeek Harness untrusted search path code execution

CVE-2026-95182 · Severity: critical · CVSS 10 · Published 2026-10-16

DeepSeek Harness (dsh) is a developer tool that executes AI agent workflows and integrates with external services. An untrusted search path vulnerability allows attackers to inject malicious executable files that get run with full tool and credential access, enabling persistent infrastructure takeover without user awareness. An attacker can embed hidden instructions in web content, emails, or code that the AI model reads, causing it to unknowingly plant a fake executable that hijacks all subsequent server connections.

Vendors: DeepSeek.

Quick answers

Is CVE-2026-95182 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-95182 being exploited in the wild.
How severe is CVE-2026-95182?
Critical severity, with a CVSS score of 10, published 16 October 2026.
Which products and versions are affected?
DeepSeek Dsh-Mcp-Client affected in packages/mcp/mcp-client/src/transport.ts; DeepSeek Dsh-Sandbox-Local also vulnerable in bwrap/sandbox-exec executors; DeepSeek Harness (dsh) commit c291e796 (2026-09) and earlier.
Is there a fix for CVE-2026-95182?
No fixed version is recorded yet. Check the vendor's advisory in the references below. If you run it, Junglewise drafts a mitigation that fits your setup and applies it once you approve, with senior security engineers behind it.
How is CVE-2026-95182 attacked?
It is attacked over the network, in a software library.
Am I affected by CVE-2026-95182?
Check whether you run DeepSeek Dsh-Mcp-Client, DeepSeek Dsh-Sandbox-Local or DeepSeek Harness (dsh) at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

Technical details

An untrusted search path vulnerability (CWE-426) in the mcp-client transport component spawns MCP servers using unqualified command names that resolve via the host PATH, which includes the writable node_modules/.bin directory. In pnpm workspace deployments, an attacker-controlled model write can place a forged executable at node_modules/.bin/{command} that gets invoked repeatedly on spawn/reconnect, establishing a persistent process with inherited credentials and full MCP tool permissions. The attack requires no user interaction: external content injection prompts the model to write the malicious file, and subsequent infrastructure processes execute it invisibly with unrestricted network egress.

Affected products

  • DeepSeek Harness (dsh) commit c291e796 (2026-09) and earlier
  • DeepSeek dsh-mcp-client affected in packages/mcp/mcp-client/src/transport.ts
  • DeepSeek dsh-sandbox-local also vulnerable in bwrap/sandbox-exec executors

Timeline

  • 2026-10-16: disclosed: CVE-2026-95182 published on NVD
  • 2026-10-14: other: PoC disclosed on GitHub gist

References

This page as Markdown. Free to reuse under CC BY 4.0 with a link back.