DeepSeek Harness (dsh) at dsh-v0.2.2-rc.2 ships with a SAFETY.md that says it has not undergone a security audit and must not be treated as secure or production-ready, and the source backs that up. Shipped profiles start sessions in workspace-write with approvals on ask, not read-only. The file sandbox governs file effects only: bwrap shares the host network namespace, Seatbelt is allow-default with only file writes denied, and Landlock grants read on the whole filesystem, so a sandboxed command can read $DSH_HOME/.credentials.yaml and reach the network. That lack of network isolation is what made CVE-2026-82533 (CVSS 4.0 9.4, CVSS 3.1 9.6) reachable from inside the sandbox; the auth hole was fixed in tag dsh-v0.1.2-alpha.1, first on npm as 0.1.2-alpha.2. On the official DeepSeek route, every request carries dsh_session_log (up to 8 MiB) and dsh_plugin_packages by default, web_search goes to the DeepSeek Search API, and feedback telemetry connects directly to a DeepSeek collector past your proxy. Routing to an internal model through llm-pi-ai drops both request fields. Installed plugins run in-process outside the sandbox, and the plugin registry falls back to registry.npmmirror.com unless you set fallbackRegistries to an empty list. Start with one disposable VM, the lockdown patch below, and an egress firewall that allows only your inference endpoint and npm mirror.
DeepSeek Harness ships its own verdict. SAFETY.md at tag dsh-v0.2.2-rc.2 opens: "DeepSeek Harness is experimental developer-preview software. It has not undergone a security audit and must not be treated as secure or production-ready." Any deepseek harness security review that ends at "the vendor says it is not production-ready" is half done, though, because a platform team still has to decide whether developers can run it, and that decision rests on the source, not the disclaimer.
This review is pinned to dsh-v0.2.2-rc.2, the GitHub prerelease tagged on 2026-11-11 (commit ddac967) and published to npm as @deepseek-ai/dsh 0.2.2-rc.2 on 2026-11-12. Every dsh release to date is a prerelease. Everything below comes from the repository at that tag, its release notes, the npm registry record and the CVE record. Nothing here was installed or run; where the source leaves a question open, we say so instead of guessing.
The short version for a regulated reviewer: shipped profiles start in workspace-write, not read-only; the sandbox has no network vocabulary; the DeepSeek route attaches your session log to model requests by default; plugins are host code. None of that is hidden. It is all in the docs, which is the good news.
What data DeepSeek Harness sends by default
The egress inventory is the core of the review. The table lists seven paths a default install can use, some only once a user signs in or installs a plugin, and two more that exist in the codebase but are not mounted by default for Web and CLI users.
Rows 1 and 2 deserve the most attention. The session-log-deepseek README says "Shipped profiles mount the plugin, so the default configuration registers the request field," and the source default is z.boolean().default(true). The 0.1.2-alpha.1 release note described session-log upload as opt-in; at rc.2 it is on, and re-enabling the switch resumes the unaccepted suffix of the log, including events recorded while it was off. The plugin inventory row is the same shape: invisible to the model, present on every official request.
Row 5 is narrower than it sounds but routes badly. FEEDBACK_ONLY releases a session-log prefix only after a user submits feedback, "regardless of model provider", and exports carry an anonymous user id from $DSH_HOME/.anonymous-user-id. The catch is in network-proxy.md: the OTLP exporter uses Node's own HTTP client, so usage telemetry "connects directly" and ignores your corporate proxy. The only supported modes are FEEDBACK_ONLY and DISABLED (FULL is rejected), and DISABLED constructs no transport at all. A non-empty DSH_TELEMETRY_DISABLED also opts the process out.
Row 4 changed in 0.1.2-alpha.1, whose note reads "Enable public WebFetch by default with SSRF protection and no per-request approval." Do not read that SSRF line as a control at rc.2. The web-fetch-http README says the provider "does not restrict destination addresses, so it reaches loopback and private-network hosts," its known limitations add cloud metadata addresses, and its URL policy rejects only non-HTTP(S) schemes, embedded credentials and over-long URLs. network-proxy.md still says a literal private address is refused, but the provider source has no such check. So the model can read intranet and metadata endpoints as well as send data out in a query string.
The finding that changes the deployment design: the deepseek-llm-api-extensions README says "The provider-neutral LLM seam and llm-pi-ai do not consume this registry." That registry is what adds rows 1 and 2 to requests. Point dsh at an internal vLLM or gateway through the llm-pi-ai route and those two fields are never attached. Rows 3 to 7 still need config and a firewall. One more detail from the rc.1 notes: web search tools hide when their required credentials are missing, so a workstation with no DeepSeek key loses row 3 by omission, but that is a side effect, not a control.
| # | Path | Default at rc.2 | Where it goes | Setting that closes it |
|---|---|---|---|---|
| 1 | dsh_session_log request field | enabled: true, up to 8 MiB per request | Official DeepSeek model API | session-log-deepseek enabled: false, or Settings > General > Upload Session Log when using the official model API |
| 2 | dsh_plugin_packages request field | enabled: true | Official DeepSeek model API (plugin names and versions) | plugin-package-inventory-deepseek enabled: false |
| 3 | web_search tool | searchProvider: deepseek-search-api | DeepSeek Search API with account token or DEEPSEEK_API_KEY | tool-web search: false (headless/SDK), firewall |
| 4 | web_fetch tool | fetch: true, no per-request approval | Any HTTP(S) host, including loopback and private-network addresses, anonymous local fetcher | tool-web fetch: false (headless/SDK), firewall |
| 5 | Feedback telemetry | FEEDBACK_ONLY | https://dsh-otel-collector.deepseeksvc.com/v1/logs, direct, bypasses proxy | DSH_TELEMETRY_MODE=DISABLED |
| 6 | DeepSeek model list refresh | On when a key is saved on the default endpoint or an account is signed in | DeepSeek API, about every ten minutes while in use | Do not save a DeepSeek key or sign in; use an internal provider |
| 7 | Plugin installs and detail pages | Registry fallback registry.npmmirror.com; detail docs from jsdelivr, unpkg, npmmirror | Public npm mirrors and CDNs | plugin-manager registry set internally, fallbackRegistries: [] |
| - | CoT translation | Experimental bundle, disabled in shipped profiles | Bing by default once enabled, Google available | Leave unmounted |
| - | Product analytics | Desktop profile only | DSH_PRODUCT_ANALYTICS_OTLP_URL | Not mounted by Web ("ordinary Web does not mount this service") |
What the DeepSeek Harness sandbox does and does not block
Start with a correction, because the package docs point the other way. The sandbox-policy package defaults to read-only, but no shipped profile uses that default. The base bundle sets mode: !!js process.env.DSH_PERMISSION_MODE ?? 'workspace-write', and the approval row resolves to ask unless the mode is danger-full-access, which pairs with never. The headless bundle repeats the workspace-write default. A fresh session can write to the workspace and asks before anything the policy flags.
Then the scope. The sandbox-policy README's Known Limitations section: "SandboxMode governs file effects; network and process policy are outside its vocabulary, so no knob here restricts them." The backend profiles in sandbox-local show exactly what that means on each OS.
Two properties are well built. The sandbox fails closed: confine() rejects with SANDBOX_UNAVAILABLE when no runner works, rather than silently running unconfined. And each execution reports full or partial enforcement, so you can log which one a workstation got. Be wary of runnerCommand, which the README describes as an operator assertion that skips the probes.
The consequence the project states itself is in the credentials-local README: "Only your OS user can read the file, but agent tool processes run as that same user, so this store cannot isolate secrets from the agent." The file is $DSH_HOME/.credentials.yaml. It holds API keys and the Web cookie signing secret, and the OS-keychain provider is deferred. Combine read on / with an open network namespace and a prompt-injected command can read the key and send it anywhere. Lookup precedence also matters for reviewers: stored file, then launch environment, then the project's dotenv file in the working directory, then the harness-home dotenv file.
The docs and release notes name several more paths that sit outside the shell sandbox entirely. Document parsing runs in the DSH host process (rc.2 note). Web user terminals run with system-user permissions independent of the agent's sandbox mode (0.1.6-alpha.2). Windows sandbox deletions could escape the authorized directory until 0.1.7-alpha.1. Plugins, covered below, run in-process.
Last, additionalWritableRoots (default []). The README warns that granting the whole Git directory "also permits changes to hooks and configuration that a later host Git invocation can execute outside the sandbox." That is the same class of defect we covered in coding agents that execute auto-loaded files from an untrusted repo: the write happens inside the box, the execution happens later on the host. For how Landlock and Seatbelt confine processes, and why a file-effect sandbox is not a network boundary, see sandboxing coding agents locally without a VM.
| Backend | Platform | What it confines | What it leaves open | Enforcement |
|---|---|---|---|---|
| bwrap | Linux (first choice) | --ro-bind / /, private PID namespace (--unshare-pid), --die-with-parent | No --unshare-net: host network namespace shared; all reads allowed | Full |
| Landlock | Linux (fallback) | Writes limited to /dev/null, /tmp and workspace roots | Read on /; network untouched | Full, partial on older ABIs |
| Seatbelt | macOS | (deny file-write*) with write allow-lists | (allow default): network and reads allowed; depends on deprecated sandbox-exec | Full |
| Windows ACL | Windows | Restricted-token runner for writes | Reads unconfined; extra directory grants apply to file tools only, not PowerShell, terminals or PTC | Partial |
CVE-2026-82533: Host header authentication bypass
The missing network isolation has already produced one critical CVE. CVE-2026-82533, titled "DeepSeek Harness < 0.1.2-alpha.1 Authentication Bypass via Host Header Spoofing", is scored 9.4 under CVSS 4.0 and 9.6 under CVSS 3.1, classified CWE-807 (reliance on untrusted inputs in a security decision).
The local HTTP agent-control API accepted "a client-supplied loopback Host header in place of validating the actual TCP connection origin." Because the sandbox does not isolate the network, a confined tool process could reach that API on loopback without any port exposure, and per the record "use it to escape its own OS sandbox, escalate to unconfined execution, and disable the approval prompt." With the port exposed through a tunnel, SSH forward or reverse proxy, a remote attacker could create sessions, execute commands and exfiltrate stored transcripts without credentials.
There is no security advisory in the repository itself, but the GitHub Advisory Database carries it as GHSA-8m2g-8cgm-3vcp, filed against the dependency @deepseek-ai/dsh-client-connection below 0.1.2-alpha.1 rather than @deepseek-ai/dsh, so a scanner flags it only if it walks the dependency tree. The minimum version is anything at or after 0.1.2-alpha.2 from npm; on the post date, pin 0.2.2-rc.2.
The reviewer's lesson is about what the fix did not change. The fix replaced Host-header trust with token-and-cookie authentication. The sandbox that let a confined process open a socket to the control plane is exactly as open to the network at rc.2 as it was in August. The next local service a dsh process can reach, yours or its own, gets the same exposure.
| Date | Event |
|---|---|
| 2026-08-25 | Public date in the CVE record; patch commit 3e24087, "fix(web): authenticate the browser Host API" |
| 2026-08-27 | First GitHub tag with the fix: dsh-v0.1.2-alpha.1 |
| 2026-08-30 | First npm version with the fix: 0.1.2-alpha.2 (alpha.1 was never published to npm) |
| 2026-09-08 | CVE record published by the assigning CNA |
Web UI authentication and exposing it on a network
At rc.2 the connection README says "every Host RPC method and WebSocket stream requires an authenticated browser or API Peer; there is no method-specific loopback tier." Loopback is no longer a trust signal.
The flow: each process mints a random launch token, accepted only on GET / (exchanged for a signed cookie and a redirect) and GET /connect for API callers. There is no Authorization-header token. The cookie has an absolute expiry of 30 days by default (cookieMaxAgeDays), is bound to hostname and port, host-only, Path=/, HttpOnly, SameSite=Strict, and Secure only when the listener serves TLS. There is no logout operation; revoking every session means deleting the owner credential record and restarting. The web-app README adds that the credential never appears in the model context or in DSH_WEB_URL, which closes the obvious prompt-injection route to the token.
The Host and Origin fence admits loopback, the bind IP, or a --trusted-host entry. The README is careful: "These checks defend DNS rebinding and cross-site browser requests; they never establish identity." A fence failure returns 403, a missing credential 401.
For exposure, the defaults are sound and the flags are easy to misread:
127.0.0.1:3080, and --host accepts one concrete IP literal; wildcard addresses are rejected.--public-url is display only and does not admit anything.--tls-cert and --tls-key, or a TLS proxy on the same host.For a workstation pilot, keep the Web UI on loopback and do not expose it at all.
Plugin security: installed plugins run outside the sandbox
The plugin-manager README is unambiguous: "installed Host code executes in-process outside the workspace sandbox." A plugin is an npm or Git package running with the dsh process's rights, which are your OS user's rights. Everything said above about the sandbox is irrelevant to plugin code.
The controls that exist are approval controls. The plugin_manager tool is enabled in the Standard and Creator Web presets, and "every tool action requires danger-full-access or approval for that call." Installs run through pnpm; dependency build scripts are blocked until a user clicks "Allow these scripts and retry", a separate approval. The ui-plugin-manager README flags a subtle one: pnpm's minimumReleaseAge can install a different version than the one the user inspected, and the UI names both. Read that dialog before approving.
The registry chain is the config fix. registry defaults to whatever pnpm is configured with, and fallbackRegistries defaults to ['https://registry.npmmirror.com/'], asked whenever the primary is unreachable or lacks the package. For a regulated network, set registry to your internal mirror and fallbackRegistries: [], otherwise an outage on your mirror silently sends installs to a public one.
Custom plugin indexes are an experimental, default-disabled bundle, and "DSH ships no index source." The trust model in plugin-market.md is stated plainly: "An index chooses each plugin's title, description, icon, author, and source; DSH checks the file format but not those claims." An http: index can be rewritten in transit, install specs included. The install confirmation shows the npm package and registry, or the Git repository, commit and directory, and Git sources must pin a 40-hex commit. Since rc.2, indexes can also display GitHub stars and weekly npm downloads; neither is a trust signal.
The operable control is an internal https index listing only bundles your team has reviewed, each Git source pinned to a full commit, backed by the internal npm mirror. That is the same supply-chain shape as MCP servers, and the review process in auditing MCP servers for tool poisoning and supply-chain attacks carries over.
Pointing dsh at an internal vLLM or model gateway
This is the step that turns rows 1 and 2 of the egress table off by construction, so it belongs in a security review and not only in a setup guide.
dsh's multi-provider route is the llm-pi-ai package, built on Pi's provider library (@earendil-works/pi-ai ^1.3.0); we compared Pi itself against Claude Code and OpenCode in Pi coding agent vs Claude Code vs OpenCode. You configure it either through the Custom model API card in the Web UI or directly under the llm-pi-ai row in $DSH_HOME/profiles/<profile>/cordis.patch.yml (the card writes the same file; the web profile's is $DSH_HOME/profiles/web/cordis.patch.yml). Supported protocols are openai-completions, openai-responses, anthropic-messages and google-generative-ai.
Three details catch internal deployments:
compat.supportsDeveloperRole: false and compat.maxTokensField: max_tokens; otherwise the output cap is sent as max_completion_tokens.deepseek-flash on the deepseek-official provider. Select the internal model before the first session; per providers.md, "Selecting a model also makes it the default for new sessions."Model discovery calls GET /models on your endpoint, so allow it at the firewall. The server side of this, a model served inside the perimeter with no outbound path of its own, is covered in deploying vLLM air-gapped.
Lockdown config and a one-machine pilot checklist
Here is a profile patch built only from row ids and keys present at dsh-v0.2.2-rc.2. It disables both DeepSeek request fields and telemetry, turns off web tools for headless and SDK profiles, pins plugin installs to an internal mirror with no public fallback, and routes the model to an internal endpoint. Hostnames and the model id are placeholders. A Cordis config override replaces that row's complete config, so each block below is the whole config for its row.
# $DSH_HOME/profiles/<profile>/cordis.patch.yml (dsh 0.2.2-rc.2)
# Each override replaces that row's complete config.
- id: session-telemetry-otel
config:
mode: DISABLED
- id: session-log-deepseek
config:
enabled: false
- id: plugin-package-inventory-deepseek
config:
enabled: false
# headless/SDK profiles; Web composes web tools per agent preset
- id: tool-web
config:
search: false
fetch: false
- id: plugin-manager
config:
registry: https://npm.internal.example/
fallbackRegistries: []
- id: llm-pi-ai
config:
providers:
internal-llm:
api: openai-completions
baseURL: https://inference.internal.example/v1
apiKeyEnv: INTERNAL_LLM_KEY # placeholder value is fine for a keyless server
compat:
supportsDeveloperRole: false
maxTokensField: max_tokens
models:
- id: your-model-idThe tool-web block is scoped deliberately. In the Web profile the host tool-web row is disabled and each agent preset mounts its own copy (the Standard preset sets fetch: true), and we have not confirmed that a profile patch reaches those preset rows. For Web users, treat web search and fetch as open until the firewall says otherwise.
The sandbox mode and telemetry are set by environment, which feeds both the sandbox-policy and approval rows in the shipped bundle:
export DSH_PERMISSION_MODE=read-only export DSH_TELEMETRY_MODE=DISABLED npx @deepseek-ai/dsh@0.2.2-rc.2 web --no-open
read-only keeps approval on ask. The Web UI stays on its default loopback bind, and --no-open skips the browser launch. For headless runs, dsh --profile headless --permission-mode <preset> --approval-policy ask gives the same posture per invocation.
None of this replaces the boundary SAFETY.md asks for: "Prefer a disposable virtual machine, container, or dedicated environment." The firewall around that VM is the only control that covers three paths config cannot reach. Telemetry, if anything re-enables it, bypasses the proxy. Workflow workers and Node ptc-runtime processes "receive no proxy settings," so model-authored scripts go direct. And sandboxed commands have no network isolation at all. NO_PROXY is not an egress control either: loopback always goes direct, and CIDR entries in it do not work. Two proxy details count in dsh's favour: a project's own dotenv file cannot set proxy variables ("DSH refuses to start rather than let a repository decide where your traffic goes"), and SOCKS is unsupported, so there is no surprise transport. A proxy password embedded in the URL is readable by every command the model writes, so authenticate the VM to the proxy another way.
This is where on-premise deployment changes the answer rather than decorating it. On DeepSeek's route, a default install sends the session log and plugin inventory with every request and searches through DeepSeek's API; for a bank, insurer or pharma team that is a data-transfer question before it is a security one. On an internal model through llm-pi-ai, those fields never exist, and the remaining paths are a short allow-list: the inference endpoint and the npm mirror. That is a policy a regulated security team can write, enforce and audit. The same container-plus-egress pattern works for other terminal harnesses; we walked through it for OpenCode in offline, air-gapped mode, and the wider tooling landscape is on the AI development tools pillar.
The pilot checklist, for one machine:
@deepseek-ai/dsh pinned to 0.2.2-rc.2 by exact versionDSH_PERMISSION_MODE=read-only, DSH_TELEMETRY_MODE=DISABLED$DSH_HOME/.credentials.yaml holds only the internal model key, no DeepSeek key, no personal tokensfull or partial) logged; Windows hosts treated as partialThis week, pick one developer VM, pin @deepseek-ai/dsh at 0.2.2-rc.2, apply the patch, deny all egress except the inference endpoint and registry, and read the firewall's deny log after the first day. Every deny entry maps to a row in the egress table above, and any that does not is the finding your review is missing.
FAQ
Quick answers to the questions this post tends to raise.



