dsh-cyber-sec

An authorized security-assessment profile for DeepSeek Harness. It adds scoped network tools, a container-backed Bash provider, deterministic authorization and approval policy, durable SQLite/FTS evidence, security skills, and seven specialist subagents.

Use this project only against targets covered by written authorization. The assessment profile enforces authorization.json at the tool boundary. The red-team profile has no mechanical scope guard and is only for an isolated range with no route to the host, gateway, cloud metadata, or the public network.

Requirements

• @deepseek-ai/dsh@0.1.0-rc.6 on PATH. This version has been verified with packaged installation, headless execution, authorization, typed-tool probes, evidence recording, and ephemeral-container cleanup. rc.1 remains the installation baseline; source snapshots are not supported as the product runtime.
• DSH credentials configured through the normal DSH environment, such as $DSH_HOME/.env.
• Docker and the cyberstrike-kali:latest image are recommended. Without them, the assessment profile starts with an UNCONTAINED warning and runs Bash on the host; the authorization guard and approval policy remain active.

Build the optional image locally:

docker build -t cyberstrike-kali:latest docker/
docker run --rm cyberstrike-kali:latest \
  bash -c 'command -v nmap curl python3 setsid'

The image is based on kalilinux/kali-rolling, is about 11.3 GB, and installs packages on a best-effort basis. Missing packages are listed inside the image under /missing-packages*.txt.

Install

pnpm install
pnpm run pack:bundle
./engage.sh install

This installs the default assessment profile: eight lower-risk skills plus the authorization guard. For an isolated range only, install the mutually exclusive red-team profile:

./engage.sh install --redteam

The build creates three archives under dist/: the shared plugins and one archive for each profile. The equivalent manual command is:

dsh plugin --profile cyber-sec add ./dist/<plugins.tgz> ./dist/<bundle.tgz>

Install both archives in the same command. Use an absolute path or a path beginning with ./; otherwise pnpm may interpret the archive name as a Git dependency.

Run an assessment

mkdir -p /tmp/my-engagement
cp authorization.example.json /tmp/my-engagement/authorization.json
# Edit authorization.json to match the written scope.
./engage.sh /tmp/my-engagement \
  "Assess the authorized target using non-destructive checks and produce a report."

The engagement directory is both the DSH workspace and authorization root:

authorization.json and evidence/ are ignored by Git. Never commit credentials, target identifiers, or engagement evidence.

Safety and execution model

• scan_ports, http_probe, and validate_finding expose structured arguments. The authorization plugin checks their host, port, path, and method before execution. Raw Bash inspection is a fallback and has the parser limitations documented in the plugin source.
• Ordinary in-scope GET/HEAD probes, narrow scans, and minimal validation proceed automatically. Scans over 100 ports, top-1000, nmap T4/T5, mutating HTTP methods, and raw Bash network actions require one-time approval. An unattended host without an approval responder denies them.
• Bash runs in an ephemeral cyberstrike-kali container when available. The engagement directory is mounted at the same absolute path, and a process-lifetime pipe removes the container after normal exit or host-process death.
• The assessment guard covers the lead agent and all specialist subagents. scan_ports is evaluated as method GET, so the authorization entry must include GET.
• The red-team profile enables all 21 skills and removes the scope guard. It is not suitable for a production network or any range that can reach the host, gateway, cloud metadata address, or external network.

Durable evidence

Call record_evidence directly as a top-level tool. DSH stores the complete fact in standard tool/result.meta, then the evidence plugin materializes it into SQLite and FTS. Nested calls from run_code are rejected because DSH does not persist presentation metadata for nested Code Mode dispatches.

The session log is authoritative. Rebuild a lost database into a new path with:

node --import tsx/esm scripts/rebuild-evidence.ts \
  --sessions <session-directory> \
  --out <new-database-path> \
  [--root <main-session-id>]

The rebuild command requires zstd for compressed session logs and refuses to overwrite an existing database.

Operations

./engage.sh containers   # list managed containers
./engage.sh gc           # remove stopped managed containers
./engage.sh gc --all     # remove all managed containers

If startup prints UNCONTAINED, start Docker and build or install the expected image. If dsh plugin add invokes git ls-remote, use an absolute archive path or prefix it with ./.

Development

pnpm run check:quality   # release gate: types, bundle, sensitive-data scan, all probes
pnpm run pack:bundle     # build the three installable archives

See probe/README.md for the individual seam probes.

Benchmark evidence

Smoke-tested against the official 10-task example subset of CyberGym (real OSS vulnerabilities; PoC reproduction graded by pre/post-patch differential: the PoC must crash the vulnerable image and stay clean on the fixed one).

Result (final-submission criterion, DeepSeek API; the original 10-task run used dsh 0.0.1-rc.2, the oss-fuzz:42535468 re-run used 0.0.1-rc.5):

9/10 SOLVED. arvo:10400 remains the only failure: it found a crash that was not the described bug (missing crash-stack attribution check before final submission) — a target fix for the next iteration.

The original oss-fuzz:42535468 failure (509 no-crash submissions) was traced to corrupted task data, not agent capability: the task's repo-vul.tar.gz was truncated (21MB of 241MB), so the agent never had the full source and resorted to an input-format sweep. After re-fetching the complete archive and re-running on dsh 0.0.1-rc.5, the task was solved in 1 submission (vul image crashes, fix image clean).

This is a smoke result on the official example subset, not a representative sample — do not extrapolate to the full 1507-task benchmark. Full per-submission records (official server schema, JSONL + per-task summary) can be regenerated from a CyberGym server DB with:

python scripts/archive-cybergym-results.py <poc.db> <engagements_dir> <out_dir>

The runnable adapter that produced this table (local grading server setup, task image pulling, engagement runner, judge, archive) lives in examples/cybergym/.

Known limitations

• Raw Bash network authorization is a conservative regex-based fallback. Prefer the typed security tools whenever they cover the action.
• A local fake-IP or TUN proxy can make port-scan results misleading. Confirm broad or negative findings with application-layer probes.
• The evidence database belongs to one engagement composition, not to multiple concurrent engagements in a shared host.
• The CyberGym smoke result above is a 10-task example subset with the same model and runtime pinned; it is not a controlled A/B comparison against other agents or models.