Playbooks · Coding tasks
Check DeepSeek's dsh tool is really running the version with the sandbox-escape fix
Written October 30, 2026
For anyone running DeepSeek's dsh tool — dsh is DeepSeek's harness, the program that runs the agent. Find the copy of the program that actually executes, and read its version: a version number on a download page is not proof of the version your machine runs, and a desktop wrapper can carry its own copy. If the fixed version is not the version that runs, an agent inside its sandbox — the restricted workspace meant to confine it — can switch its own session to danger-full-access, turning off the sandbox and its approval prompts.
The flaw, in brief
CVE-2026-82533, rated CVSS 9.4, covers versions of the tool before 0.1.2-alpha.1. The flaw was in the tool's local control interface: it trusted a client-supplied loopback Host header instead of the real connection origin, so a sandboxed agent — an agent running inside the sandbox, the restricted workspace meant to confine it — could call that interface and set its own session to danger-full-access, which turns off the sandbox and stops approval prompts. The sandbox confined file writes, not reads or network access, so the interface was reachable from inside it. The fix adds one-time-token authentication to the interface. It does not make the sandbox confine reads or network access.
The version number needs care. The fixed code shipped on GitHub as 0.1.2-alpha.1 on 27 August 2026, but that build never reached npm. The first fixed release on npm was 0.1.2-alpha.2, on 30 August 2026. If you install from npm, 0.1.2-alpha.2 is the floor — a machine reporting 0.1.2-alpha.1 from a source build and a machine reporting an older npm release are in different positions, and only the version that actually executes counts.
Steps
- Find the binary that executes. Run:
command -v dsh— this prints the path of thedshyour shell actually runs.dsh --version— this prints that binary's version.
- Find the copy inside any desktop wrapper, separately. If you launch the tool from a desktop app or wrapper, that app may bundle its own copy. Look inside the app's install folder for a bundled
dshbinary, or read the version the app itself reports, and run the bundled copy's full path with--version. Record its path and version on a separate line. Do not assume the wrapper's copy matches the command-line copy. - Compare each copy against the floor. Any copy below 0.1.2-alpha.2 installed from npm needs an upgrade. A copy at 0.1.2-alpha.1 built from GitHub carries the fix, but it is not an npm release — if you manage installs through npm, move to 0.1.2-alpha.2 or later so future updates follow the normal path.
- Upgrade from npm. The official npm package is
@deepseek-ai/dsh— that is the package DeepSeek's own README runs from npm (npx @deepseek-ai/dsh web). Run:npm install -g @deepseek-ai/dsh@0.1.2-alpha.2
npxrather than a global install, check that copy instead:npx @deepseek-ai/dsh --version. Then repeat step 1 and confirm the version now reports 0.1.2-alpha.2 or later — the path may stay the same, and that is expected. For a bundled wrapper copy, update the wrapper itself through its own update path, then repeat step 2 — upgrading the command-line copy does not upgrade the bundled one. - Repeat the pass after any wrapper or tool update. A wrapper update can replace its bundled copy with an older or newer one. The version that runs is the version you read after the update, not the version the update notes claim.
What can go wrong
- Reading the wrong copy's version. The command-line copy can be fixed while a desktop wrapper still runs an older bundled copy. The vulnerable copy is the one that executes, whichever path it lives at.
- Treating a GitHub fix as an npm fix. 0.1.2-alpha.1 fixed the code on GitHub but never reached npm. An npm-based install below 0.1.2-alpha.2 is still exposed, even though a fix exists somewhere.
- Exposing the local interface. If the tool interface is reachable from outside the machine — through a tunnel, a port forward, or a reverse proxy — a remote caller could reach the same control surface. Keep the interface local, and do not forward its port.
- Assuming the patch confines everything. The fix adds authentication to the control interface. Under the default setting, reads and network access are still not confined, so the sandbox is a write boundary, not a full isolation boundary. Treat agent work on untrusted material accordingly.
Limits
- This pass reads version numbers. It does not test whether a given copy is exploitable, and it does not audit anything else the tool runs.
- Wrapper internals differ between apps. If a wrapper gives no way to read its bundled version, that is a reason to ask its vendor, not a reason to assume it matches.
Related playbooks
- What version should I be on
- Check your Codex and Claude Code config for relay tampering: base URL, sandbox, and credentials you did not set
- Check your first DeepSeek search/fetch bill
- Check whether your session is really on Muse Wonder or Wonder Full — and which route you are paying through
Sources
- http://thehackernews.com/2026/09/deepseek-harness-flaw-let-ai-agents.html
- https://neomanex.com/news/deepseek-harness-cve-2026-82533-sandbox-escape
- https://cyber-security-in-plain-english.com/post/developers/news/deepseek-harness-sandbox-control-plane/
- https://www.llm-hacking.com/hacks/deepseek-harness-sandbox-self-disable.md/
Back to: Coding playbooks hub