Security-audit methodology skill pack plus the plugin_vet supply-chain gate: eight agent skills (secret scan, dependency audit, supply-chain review, prompt-injection review, audit orchestration, threat modeling, vuln intel, incident response) in Chinese and English editions, with an npm provider bundle that mounts the skills and registers the automated plugin_vet pre-install scanner.
Install
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:PerryLink/dsh-skill-pack-security
Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).
README
dsh-skill-pack-security
Lay the supply-chain risk out before dsh plugin add — the verdict, not a hunch, decides the install.
Eight security-audit skills plus an automated pre-install gate for DeepSeek Harness: the skills teach the methodology, plugin_vet runs the mechanical checks, and every finding points back into a skill section.
Install · Configuration · FAQ · 简体中文 · English · Español · Português · हिन्दी
dsh plugin --profile <name> add dsh-skill-pack-security
Community plugin, not affiliated with DeepSeek. Canonical repository: https://github.com/PerryLink/dsh-skill-pack-security (Gitee is a mirror).
This project has moved to dsh-security-pack: this repository stays as a pointer/backup and new work happens there; the old npm package @perrylink/dsh-skill-pack-security-provider is deprecated.
What it is
dsh-skill-pack-security is a skill pack plus a supply-chain gate. Eight security methodologies ship as SKILL.md bundles the model discovers in its session catalog and loads on demand, while plugin_vet runs the mechanical pre-install checks — license, SBOM, commit locking, lifecycle scripts and exfil domains, a five-dimension risk card — and closes with PASS / WARN / FAIL. The skills answer "how do I audit this"; the tool answers "what does the package look like before I install it"; both share one vocabulary and one set of section numbers. The root package is a private bundle manifest; the runnable package is @perrylink/dsh-skill-pack-security-provider.
owner/repo ──tarball (once)──▶ plugin_vet ──▶ license · sbom · commit-lock
lifecycle · exfil · obfuscation
data-responsibility
│
▼
risk card (5 dims x 0-100) ──▶ PASS | WARN | FAIL ──▶ gate.policy
Three-step setup
Install (pick one channel; the runnable package is the provider):
# git channel (latest main) dsh plugin --profile web add github:PerryLink/dsh-skill-pack-security # npm channel (published releases, with provenance) dsh plugin --profile web add @perrylink/dsh-skill-pack-security-provider # tarball channel: pnpm pack inside provider/, then dsh plugin --profile web add ./@perrylink-dsh-skill-pack-security-provider-<version>.tgzprepackembeds both editions —skills/(Chinese) andskills-en/(English) — into the tarball; the row that mounts is the one inprovider/cordis.patch.yml.Minimal config (defaults to the Chinese edition; one key switches it):
- insert: - id: skill-pack-security name: '@perrylink/dsh-skill-pack-security-provider' config: language: enVerify it mounts:
dsh --profile web --dump-config | grep -A3 'id: skill-pack-security'Then have the model load
security-auditwith theskilltool and callplugin_vetonce against anyowner/repo— a rendered risk card means it works.
Uninstall: dsh plugin --profile web remove @perrylink/dsh-skill-pack-security-provider (or remove the row; pure-skill copies are removed with the installer's -Uninstall / --uninstall).
Every PerryLink plugin is listed in the built-in DSH Desktop market: Market → Sources → Add source → paste https://perrylink-dsh-catalog.perrylink.workers.dev/catalog-source.json → select. Installation still requires the market's npm identity check and your confirmation.
Installing the skills by hand
You do not need the provider to use the skills. DSH's local skill provider scans four roots by rank (lower rank wins same-name conflicts within a layer):
| Rank | Root | Scope |
|---|---|---|
| 100 | <projectRoot>/.dsh/skills |
Project-scoped, travels with the repo |
| 200 | <projectRoot>/.agents/skills |
Project-scoped, shared agent directory |
| 400 | <dshHome>/skills ($DSH_HOME or ~/.dsh) |
User-scoped, DSH-only |
| 500 | <agentsHome>/skills ($DSH_AGENTS_HOME or ~/.agents) |
User-scoped, cross-agent |
Rank chain (lower wins a same-name conflict within a layer): project-dsh 100 < project-agents 200 < custom 300 < user-dsh 400 < user-agents 500. Custom rank 300 is the plugin-registered layer (such as this pack's optional provider/), not a disk root.
./scripts/install.ps1 -Target user-agents -Language zh # Target: project-dsh | project-agents | user-dsh | user-agents; Language: zh (default) | en
bash ./scripts/install.sh --target user-agents --language en
Core capabilities
- Eight skills, two editions — every skill ships with identical names and metadata in
skills/(Chinese) andskills-en/(English); install one language per root. plugin_vetgate — a zero-dependency scanner registered onctx.toolsby the optionalprovider/plugin (license / SBOM / commit locking / malicious patterns / data-responsibility review / five-dimension risk card).- Findings cite the skills — every finding points to the matching skill section (for example
supply-chain-review §1) so the agent can continue the manual audit. - Executable by a model — each skill step is a real command (
gitleaks,trivy,pnpm audit,npm view,git …) with an expected-output sample and an exit-code criterion.
The eight skills
| Skill | Purpose | When to use |
|---|---|---|
security-audit |
Five-phase audit flow: scope → inventory → risk tiering → verification → report template | Whole-repo audits, audit reports, planning |
secret-scan |
Credential audit: gitleaks/trivy usage, false-positive tiers, redacted reports, remediation order | Secret scanning, alert triage, leak reports |
dependency-audit |
Supply-chain audit: pnpm/npm audit reading, licenses, typosquat risk, lockfile drift | Dependency review, audit-report interpretation |
supply-chain-review |
Quick PR/new-dependency review: dangerous install scripts, typosquat, reproducible builds | Reviewing PRs that add dependencies |
prompt-injection-review |
Injection-surface review for agent projects: AGENTS.md, skills, tool descriptions, MCP, web | Reviewing model-context injection surfaces |
threat-model |
Design-stage threat modeling: trust boundaries, STRIDE table, attack trees, mitigations | Modeling new features, design-stage security review |
vuln-intel |
Vulnerability intelligence: NVD/CISA-KEV/GHSA/OSV lookups with verdict criteria | Given a CVE/GHSA id, checking impact and exploitation |
incident-response |
Agent-environment incident response: contain → evidence → recover → postmortem | Suspected security incidents in DSH/agent setups |
Each bundle keeps its main file ≤ 300 lines (progressive disclosure; details live in references/).
plugin_vet — the automated pre-install gate
plugin_vet is the pack's automated complement: a zero-dependency scanner registered on ctx.tools by the provider/ plugin. Point it at a GitHub owner/repo or a local package path — it downloads the tarball once (timeout + AbortSignal respected), scans within budget limits, and returns a render card.
- License scan — finds the LICENSE file and the
licensefield;NOASSERTION/UNKNOWN/SEE LICENSE IN <file>, a missing file, or a missing field is flagged; common SPDX ids are recognized. - SBOM — extracts the dependency tree with versions from the lockfile (pnpm/npm/yarn); when
osv-scannerornpm auditis detected on a local target, their vulnerability output replaces the self-computed dependency scan and the report annotates the source (builtin/osv-scanner/npm-audit). - Commit locking — install-manifest refs and workflow actions must be immutable 40-hex commit SHAs;
@tag/branch refs are flagged as mutable. - Malicious patterns — lifecycle scripts (
preinstall/install/postinstall), network-exfiltration domains, and obfuscated/encoded payloads in shipped code. - Data-responsibility review — the policy-scan dimensions as deterministic rules: ungated listeners on sensitive seams (
agent/pre-step,tools/pre-execute,session/event, …), outbound endpoints without README telemetry/privacy disclosure, description-behavior keyword coverage, and embedded instruction-override payloads in shipped text (skills, docs, prompts, tests). Every finding citesprompt-injection-reviewfor the manual deep-dive; disable per deployment withvet.dataResponsibility: false. A model-assisted review stage is the documented future upgrade. - Five-dimension risk report — license / source / dependencies / build scripts / maintenance, each 0-100, folded into an overall verdict: PASS, WARN, or FAIL.
Install gate. The verdict feeds an installation gate — gate.policy: warn (default, non-blocking) prints a warning on FAIL; gate.policy: deny returns a deny verdict. The tool does not intercept the host's dsh plugin add; wire the verdict into your installer/CI to enforce it:
- id: skill-pack-security
name: '@perrylink/dsh-skill-pack-security-provider'
config:
language: en
vet:
gate:
policy: deny # return a deny verdict for targets that fail plugin_vet (caller enforces)
Complementary to dsh-plugin-check. The official plugin validator's 36 checks verify a plugin's contract and quality (config schema, effect registration, tool JSON shape); plugin_vet verifies the supply chain of where a plugin comes from. Run both:
dsh-plugin-check (36 checks) |
plugin_vet (this repo) |
|
|---|---|---|
| Question answered | Is this plugin well-formed and contract-compliant? | Is this package safe to install? |
| Looks at | Plugin code, schema, registrations, tool contracts | LICENSE, lockfile, install refs/actions, lifecycle scripts, exfil/obfuscation, maintenance, data-responsibility (hooks scope, telemetry disclosure, description-behavior, embedded injection payloads) |
| Verdict | Pass/fail per check | PASS / WARN / FAIL + gate |
| When | Plugin development or review | Before dsh plugin add, PR review, CI supply-chain gate |
| Blocking | CI gate (non-zero on violations) | Configurable: warn (default) or deny |
Why skills, not tools
| Shape | What it does | What it cannot do |
|---|---|---|
| Tool plugin (e.g. security scanners) | Executes scans, returns findings | Interpret alerts, tier false positives, write redacted reports |
| Protocol layer | Constrains a protocol | Generalize across repos and agents |
| Skill pack (this repo) | Teaches methodology: triage, reporting, remediation order — and automates the static pre-install checks via plugin_vet |
Replace a manual audit end to end |
Installed together with a tool-type security plugin, the two compose: the tool runs the scan, the skill drives interpretation, triage, and the report. This pack combines both shapes: the skills teach the methodology, and plugin_vet runs the mechanical static subset automatically, with every finding pointing back into the skills.
The Claude Code ecosystem's 3000+ skills prove the distribution value of this shape. DSH's SKILL.md frontmatter (name, description, whenToUse) is format-compatible with CC skills; this pack uses only the common subset and its content is entirely original.
Tools and surfaces
| Surface | Kind | Notes |
|---|---|---|
plugin_vet |
tool | Pre-install supply-chain scan (license / SBOM / commit lock / malicious / risk card); findings cite skill sections |
skill-pack-security |
skill provider | Registers the pack's skills/ or skills-en/ edition on ctx.skills |
Eight SKILL.md bundles |
skills | The audit methodology, in two language editions |
| install gate | gate | vet.gate.policy: warn | deny returns a verdict; the caller (installer/CI) enforces it |
Measured data
The per-class detection rate / FPR / F1 over 38 poison samples (20 malicious, 18 benign) is published in benchmark/RESULTS.md; regenerate it with pnpm --dir provider run build && node benchmark/run.mjs (zero new dependencies). What is measured is the three malicious-pattern checks in the shipped engine — installScriptsCheck, networkExfilCheck, obfuscationCheck; only a fail verdict (the state that blocks an install under the deny policy) counts as a detection, a warn does not. Micro recall 0.850, macro F1 0.903, micro F1 0.919, zero false positives. The three missed samples and the gap against OSV/Socket are written out in RESULTS.md: these heuristics read shape only and carry no vulnerability database, so CVE intelligence belongs to the sbom check (needs osv-scanner/npm audit).
Configuration
All tunables are Schemastery Config fields (changeable from cordis.yml). provider/cordis.patch.yml documents each key inline.
Most-used keys:
| Key | Default | Meaning |
|---|---|---|
language |
zh |
Edition to publish: the Chinese skills/ or the English skills-en/; ignored when skillsDir is set |
skillsDir |
(unset) | Explicit skills root; overrides the language-derived default and must hold <skill>/SKILL.md bundles |
vet.enable |
true |
Register the plugin_vet gate tool |
vet.dataResponsibility |
true |
Run the data-responsibility review (disable per deployment) |
vet.externalScanners |
true |
Orchestrate osv-scanner/npm audit when their CLIs are present; false forces the built-in self-computed dependency scan |
vet.gate.policy |
warn |
Install gate: warn (non-blocking) or deny (block on FAIL) |
| Key | Default | Meaning |
|---|---|---|
language |
zh |
Edition to publish: the Chinese skills/ or the English skills-en/; ignored when skillsDir is set |
watch |
false |
Watch the packaged skills directory (static content, so disabled) |
skillsDir |
(unset) | Explicit skills root; overrides the language-derived default and must hold <skill>/SKILL.md bundles |
vet.enable |
true |
Register the plugin_vet gate tool |
vet.timeoutMs |
15000 |
Tarball-fetch timeout in ms |
vet.maxFiles |
800 |
Scan file cap |
vet.maxFileBytes |
262144 |
Per-file byte cap |
vet.maxExtractBytes |
67108864 |
Extraction byte cap |
vet.maxDepNodes |
600 |
Dependency-tree node cap |
vet.maxFindingsPerCheck |
12 |
Findings cap per check |
vet.dataResponsibility |
true |
Run the data-responsibility review (disable per deployment) |
vet.externalScanners |
true |
Orchestrate osv-scanner/npm audit when their CLIs are present; false forces the built-in self-computed dependency scan |
vet.userAgent |
dsh-skill-pack-security/2.2.25 (+https://github.com/PerryLink/dsh-skill-pack-security) |
Fetch user-agent |
vet.gate.policy |
warn |
Install gate: warn (non-blocking) or deny (block on FAIL) |
Example override in your profile patch:
- insert:
- id: skill-pack-security
name: '@perrylink/dsh-skill-pack-security-provider'
config:
language: zh
vet:
gate:
policy: deny
FAQ
plugin_vet downloads the target tarball once (respecting vet.timeoutMs and AbortSignal), extracts and scans it locally, and redacts secret-shaped text from the rendered report; the plugin injects no prompt sections and does not send reports anywhere. The dshWorkshop manifest declares exactly two permissions: files:read and network:fetch. The skills themselves are plain text with no executable code.
Nothing is silent. A skillsDir that is empty or nonexistent fails the provider loudly at mount; when osv-scanner/npm audit are absent, vet.externalScanners falls back to the built-in self-computed dependency scan and annotates the source in the report; with vet.enable: false the plugin_vet tool is simply not registered (the skills keep working).
No contention: the plugin registers one tool (plugin_vet) and one skill set, and the skill names have zero overlap with the official skills or known community packs (see "Interoperability with other DSH plugins"). Keep running the official dsh-plugin-check too — it covers contract and quality, plugin_vet covers supply chain; two different questions, two different rulers.
The checks are static heuristics — zero dependencies, no dynamic execution — so the cost is one tarball download plus extraction; four budget caps (vet.maxFiles 800, vet.maxFileBytes 262144, vet.maxExtractBytes 67108864, vet.maxDepNodes 600) and a 15-second timeout bound the worst case. For measured accuracy see "Measured data" (38 samples, micro F1 0.919).
Security boundaries
- Zero-dependency engine.
plugin_vetuses onlynode:builtins and relative imports. - Narrow pre-install gate. Not a general-purpose security-audit tool — deliberately complementary to scanner plugins and the official
dsh-plugin-checkcontract validator. - Non-blocking by default. The install gate is
warnunless you opt intodeny. - Original content. Format-compatible with Claude Code skills, but no copied CC skill content and no skill marketplace.
- Original engine, no third-party ports. The license scan and malicious-pattern checks are original zero-dependency implementations; the GPL-Radar / LLM-detective / Sus-PY assets were evaluated for porting but no licensed public source was found — see
THIRD_PARTY_NOTICES.md.
Permissions & data
- Permissions: the
dshWorkshopmanifest declaresfiles:readandnetwork:fetch. - Data:
plugin_vetdownloads a tarball once (timeout +AbortSignalrespected) and reports redact secret-shaped text; the plugin injects no prompt sections. - Session log: this plugin declares no session event type, so the audit trail stays on the host's
tool/call+tool/resultevents.
Known limitations
- Not a full audit tool.
plugin_vetis a narrow pre-install trust gate; it cannot replace a manual, end-to-end audit. - Static scan only. The malicious-pattern and maintenance signals are heuristics over the shipped package, not dynamic analysis.
- One edition per root. Same-name skills in one root resolve by rank, so only one language edition enters a session catalog.
Interoperability with other DSH plugins
This plugin touches exactly two seams: it registers plugin_vet on ctx.tools and the pack's edition on ctx.skills. It contends for no existing name and composes with tool-type security plugins.
- Zero skill-name collisions. The eight skill names were checked one by one against the official
.agents/skills/set (12 skills, 2026-08 snapshot) and the known community packs (dsh-write-plugin,dsh-test-plugin,dsh-plugin-dev,make-dsh-plugin,find-plugins,mainline-compat), with no overlap;dhicoc/dsh-reverse-skill(85 skills) andcyzlmh/dsh-cyber-sec(21 skills) carry no same-name skill either. Method and per-query results:docs/ecosystem-conflict-check.md(2026-08-14 snapshot);verify/verify-skill-pack.mtscheck 2 asserts it mechanically. - Complementary to tool-type and protocol-type security plugins. The existing shapes on this topic are "tool" and "protocol"; this pack is "skills plus a static gate". The tool runs the scan, the skills drive interpretation, triage and reporting — different surfaces, no contention.
- Complementary to the official
dsh-plugin-check. 36 contract checks versus a supply-chain check; the comparison table is under "Core capabilities". - The bundle inserts only its own row.
provider/cordis.patch.ymlinserts exactly oneskill-pack-securityrow and never overrides a built-in row'sconfig.
Compatibility
| Surface | Status |
|---|---|
| Harness | DeepSeek Harness dsh-v0.2.1-alpha.1 (GitHub tag, verified 2026-09-25: full gate chain + profile install smoke). npm dependency line: @deepseek-ai/dsh 0.1.5-rc.2 (peers >=0.1.2-rc.1 <0.2.0 || >=0.1.5-alpha.1 <0.2.0). The session envelope keeps its ignorable field for stored-log read compatibility only - Session.append still cannot stamp it, so audit-gate behavior is unchanged. |
| Node | ^22.19.0 || >=24.0.0 (the DeepSeek Harness runtime) |
| Platforms | All (the skills are content; the provider is a host plugin). CI runs the 25 checks on both Ubuntu and Windows |
| Model | Any (skills load on demand via the skill tool; plugin_vet is deterministic) |
Ecosystem & quality
- Gitee mirror (sync mirror; GitHub is canonical)
- dshfind plugin page (installs and ranking)
- OpenSSF Scorecard report (driven by .github/workflows/scorecard.yml)
- dsh-plugin-doctor static audit (runs the SPEC R+K criteria on every push/PR)
- DSH Market (listed)
- 1024 store channel:
npm i -g dsh1024once, thendsh1024 plugin --profile web add dsh-skill-pack-security(counts toward the deepseek1024.com install ranking) - Ecosystem knowledge base — measured data, not marketing: plugin development guide · plugin-selection data · maintenance criteria
Roadmap
dsh-skill-pack-data-engineering— data pipelines, data quality, ETL checklists (same template)dsh-skill-pack-oss-collab— PR etiquette, issue triage, maintainer workflowsdsh-skill-pack-performance— profiling methodology, benchmark criteria, regression checklists- More skills inside this pack (same pure-skill boundary):
sbom-lifecycle(SBOM generation/aging/import workflows),pen-test-review(authorized-engagement scoping and report review),compliance-audit(ASVS/NIST-CSF walkthroughs) - Keep the
plugin_vetdemo artifacts fresh (docs/demos/run-demos.mjs) and thedsh-plugin-checkcomplement table accurate as the official checker adds checks
Development
pnpm --dir provider run typecheck # tsc --noEmit
pnpm --dir provider run build # tsc --noEmitOnError
pnpm --dir provider run prepack # embeds both skill editions into the tarball
tsx verify/verify-skill-pack.mts # 25-check headless verification
node benchmark/run.mjs # poison-sample regression set (build first)
pnpm run check:readmes # five-language README consistency
The hero diagram and the measured-data chart are generated programmatically by scripts/gen-readme-assets.py (Python + Pillow) and are safe to re-run: python scripts/gen-readme-assets.py.
Verification: 25 headless checks
verify/verify-skill-pack.mts imports the official dsh-skill-filesystem parser, the real skill tool, and the real tools runtime from a local deepseek-harness checkout and asserts 25 checks over both language editions:
- Layout: both editions present, 8 directory bundles each, no stray flat skills, frontmatter
namematches directory, ≤ 300 lines,references/wired,metadata.versionsynced to theVERSIONfile - No name conflicts with the official
.agents/skills/skills (derived from the checkout at run time) or known community skill packs 3–6. Per edition (Chineseskills/, Englishskills-en/): registry discovery through the official provider, fullctx.skills.get()loads, the realskilltool returning<skill_content>(unknown/invalid names rejected), and the session catalog containingname+descriptiononly —whenToUsestays out of the model catalog (official design) - 13 bad-frontmatter fixtures exercise the official fail-closed rules (missing fields, legacy camel-case keys, non-boolean values, non-kebab names, nested dirs, name mismatch); flat-file skills load and nested
**/SKILL.mdis not discovered - The optional provider plugin mounts the Chinese and the English edition via
ctx.effect(), disposes cleanly, and rejects misconfiguration (empty or nonexistentskillsDir) 9–15. Self-hardening checks: zh↔en structural parity, references wiring (no dangling/orphan files), provider version sync, documented skill-root ranks vs the official constants, POSIX-portablegrep -Epatterns, secret self-check, UTF-8-safe release checklist 16–19.plugin_vetthrough the real tools runtime: it registers onctx.tools; the compliant fixture passes; the no-license fixture fails and citesdependency-audit §3; the malicious postinstall fixture fails (scripts/exfil/obfuscation, citingsupply-chain-review §1); the gate returns a deny verdict underpolicy: deny(the caller enforces it) - The scan engine is zero-dependency (
node:builtins and relative imports only) - Report redaction keeps secret-shaped text out of rendered output
# local: auto-resolves the harness checkout beside the pack, or point it explicitly
$env:DSH_HARNESS_CHECKOUT = 'D:\deepseek-harness'
& D:\deepseek-harness\node_modules\.bin\tsx.CMD verify\verify-skill-pack.mts
# All 25 checks passed for dsh-skill-pack-security.
The same 25 checks run on GitHub on every push via .github/workflows/verify.yml — on Ubuntu and Windows — plus an install.sh/install.ps1 exercise and a standalone provider build/pack smoke that asserts the tarball carries both embedded editions and the bundle patch.
Layout
| Path | What it is |
|---|---|
skills/<name>/SKILL.md |
The eight skills (Chinese edition); frontmatter follows the official dsh-skill-filesystem contract |
skills-en/<name>/SKILL.md |
The eight skills (English edition); same names and metadata as the Chinese edition |
skills/<name>/references/ |
Progressive-disclosure detail: command matrices, triage tables, templates |
scripts/install.ps1 |
One-command Windows installer for all four roots (both language editions); records a manifest, supports -Uninstall/-DryRun/-Force |
scripts/install.sh |
The POSIX equivalent (--uninstall/--dry-run/--force) |
provider/ |
npm-installable provider bundle (declares dsh.bundle; embeds both editions in pack/ via prepack; language: zh|en); registers the skills provider AND the plugin_vet gate tool via ctx.effect(), fails loud on a bad skillsDir |
provider/src/vet/ |
The zero-dependency plugin_vet scan engine (license / SBOM / commit lock / malicious patterns / risk report) |
package.json |
Root bundle manifest: declares dsh.bundle.patch (→ provider/cordis.patch.yml) and dshWorkshop intake facts |
verify/verify-skill-pack.mts |
Headless verification against the official parser, the real skill tool, and the real tools runtime — 25 checks across both editions |
VERSION |
Single version source; every SKILL.md metadata.version and provider/package.json must match it (CI-enforced) |
benchmark/ |
The poison-sample regression set (dataset/poison.json, run.mjs, RESULTS.md) |
docs/ |
Ecosystem conflict check, release checklist, improvement plans, and the plugin_vet demos (docs/demos/) |
CHANGELOG.md / SECURITY.md / CONTRIBUTING.md |
Release history, vulnerability reporting policy, and contribution/verification rules |
.github/workflows/verify.yml |
CI: 25-check verification + installer exercise + provider build/pack smoke (Ubuntu and Windows) |
.github/dependabot.yml |
Weekly dependency updates for the provider and GitHub Actions |
LICENSE |
Apache License 2.0 |
THIRD_PARTY_NOTICES.md |
Third-party posture: zero-dependency engine, evaluated-but-not-ported assets, peer dependency licenses |
Family & thanks
This project is part of the DeepSeek Harness plugin family from PerryLink (all Apache-2.0). If it is useful, a star helps the next person find it — nothing is gated behind it.
- @PerryLink — author and maintainer: the eight skills in both language editions, the installers, the verification suite, the provider bundle, CI, and the documentation.
License
LICENSE (Apache License 2.0) © 2026 dsh-skill-pack-security contributors
Links
More in this category
kenryu42/cc-safety-net★ 2025
Pre-execution guard that blocks destructive Git and filesystem commands and access to secrets such as SSH keys and .env files before a tool call runs.
ADWMC/helm-d#helmd★ 1266
Security-analysis plugin for DeepSeek Harness: 33 tools spanning six security domains, evidence/reporting and toolbox, plus the case lifecycle, routing and tool ledger; a session-bound case workflow with an evidence chain, E-id validation and reproducible findings; 362 on-demand references; three hot-swappable persona profiles; and a four-tab workbench (armor / case / intercept / about) with case-state recovery after context compaction.
toby-bridges/api-relay-audit★ 914
Runs local security audits of AI API relays and LLM proxies from DeepSeek Harness, producing Markdown reports for prompt injection, model substitution signals, tool-call rewriting, error leakage, stream integrity, and profile-gated Web3 risks.
SeaOf0/dsh-redteam-model★ 877
Authorized-security DSH collection: nine work modes (redteam coordinator, pentest, code audit, binary analysis, attack-defense, AV evasion, incident response, cloud security, CTF solving) and fifteen runtime plugins, managed from a settings page with one-click deploy, install, update and uninstall.
howmp/dsh-pentest★ 621
Authorized pentest mode for DeepSeek Harness — exploration chain, assets and findings with a Web view.
PerryLink/dsh-auto-review★ 248
Second-model auto-review on the approval answerer chain: a read-only reviewer subagent returns structured allow/deny verdicts with reasons, fail-closed by default.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.