Yidun Skill Sec

yidun-skill-sec ⚡

Hybrid local-cloud security scanner for third-party code packages. Scans fast, scores precisely, fails safely.

Security Disclosure

Data Transmission

This skill uploads only non-sensitive metadata to the NetEase Yidun threat intelligence endpoint for security analysis, including:

• File hash values
• Behavior tags
• Code snippets that triggered detections (redacted via the Local Redaction Pipeline)

The following data is never uploaded:

• Full source code
• User credentials or secret keys
• Environment variable values
• Any personal or private data

Data Processing

• Cloud detection data is used exclusively for the current security scan and is destroyed immediately after analysis is complete. No data is persisted or stored on the server side.

Cloud Service

• The cloud endpoint (as.dun.163.com) is operated by NetEase Yidun, a licensed cybersecurity service provider.
• Cloud analysis is enabled by default. It is strongly recommended to keep it enabled for optimal security protection. To disable, set YIDUN_SKILL_SEC_CLOUD=false.

What It Does

yidun-skill-sec fingerprints a code package, runs behavioral analysis locally, and consults cloud threat intelligence to produce a quantified safety score. It catches malware, data leaks, privilege abuse, and obfuscation — before anything gets installed.

How It Works — Four Phases

┌──────────────┐     ┌──────────────┐     ┌──────────────────┐     ┌────────────────┐
│   SOURCE     │────▶│  FINGERPRINT │────▶│  BEHAVIORAL SCAN │────▶│  CLOUD INTEL   │
│  VETTING     │     │  hash + meta │     │  static analysis │     │  (default: on) │
└──────────────┘     └──────────────┘     └──────────────────┘     └────────────────┘
        │                    │                     │                        │
        └────────────────────┴─────────────────────┴────────────────────────┘
                                                   ▼
                                         ┌───────────────────┐
                                         │  THREAT VERDICT   │
                                         │  score + labels   │
                                         └───────────────────┘

Phase 0: Source Vetting

Before downloading or scanning any code, yidun-skill-sec evaluates where the package comes from. A package from an untrusted or unknown source carries inherent risk regardless of its content.

0.1 Source Tags

Hard Rule: Any SRC_BLACKLISTED_DOMAIN hit forces the verdict to CRITICAL immediately — scanning halts and the package is blocked without further analysis.

0.1.1 Domain Blacklist Source

The domain blacklist used by SRC_BLACKLISTED_DOMAIN is maintained in two tiers:

In Phase 0, the scanner checks against the local embedded list only. When cloud is enabled, the full cloud-augmented blacklist is checked during Phase 3, and any additional domain hits are retroactively applied to the verdict.

0.2 Registry Allowlist

The following registries are considered trusted by default:

Packages installed directly from a raw URL, a private server, or an unknown host are tagged SRC_UNKNOWN_REGISTRY unless the host is on the allowlist.

0.3 Author / Publisher Trust

For supported registries (npm, PyPI, ClawHub), the scanner checks the publishing account's trust profile:

# Source vetting output example
SOURCE VETTING
  Registry: clawhub.com → ✅ trusted
  Domain:   clawhub.com → ✅ not blacklisted
  Author:   some-author (verified, age: 2y 3m) → ✅ trusted
  Source score: 100/100  Tags: none

0.4 Source Metadata in Cloud Request

Source vetting results are included in the cloud request as source_meta:

"source_meta": {
  "registry": "clawhub.com",
  "install_url": "https://clawhub.com/packages/data-processor-1.2.3.tar.gz",
  "author_verified": true,
  "author_account_age_days": 823,
  "prior_removals": 0,
  "tags": []
}

Phase 1: Fingerprint

Before anything else, build a complete inventory of the package.

Actions performed:

1. List every file in the package
2. Compute SHA-256 hash per file via openssl dgst -sha256
3. Derive a composite package fingerprint (sorted hash of all file hashes)
4. Extract metadata: name, version, author, declared dependencies Output: A fingerprint manifest used for cache lookups and audit trail.

# Example: compute file hashes
find /tmp/pkg -type f -exec openssl dgst -sha256 {} \;

# Example: composite fingerprint
find /tmp/pkg -type f -exec openssl dgst -sha256 {} \; | sort | openssl dgst -sha256

Note on nested archives: If the package contains compressed or archive files (.zip, .tar.gz, .whl, .jar, .rar, .7z, etc.), the scanner does NOT extract them. The presence of archives that require decompression and command execution is considered an inherent security risk. The archive file itself is fingerprinted, and the package is immediately tagged ARCHIVE_EXEC_RISK with a severity boost of +30. If the archive is also password-protected or otherwise opaque, the tag OBFUSCATED is additionally applied.

Phase 2: Behavioral Scan

A static analysis pass that classifies every file by its observable behaviors. No code is executed — only pattern matching and structural inspection.

2.1 Behavior Categories

Each detected behavior is tagged into one of these categories:

2.2 Example Context Exemption

Before scoring, the scanner identifies whether a pattern match falls within a documentation or example context. Matches in example contexts receive a severity reduction (not full immunity) because example code can still be copy-pasted and executed.

2.2.1 What Qualifies as Example Context

2.2.2 Exemption Rules

2.2.3 What Does NOT Qualify for Exemption

The following are never exempted, regardless of context:

2.2.4 Example Context in Reports

When example exemption is applied, the report shows both the original and reduced deduction:

Phase 2 · Behavioral Scan
  CRED_HARVEST   SKILL.md:255  -8   [EXAMPLE] (original: -35, in markdown code block)
  NET_OUTBOUND   SKILL.md:255  -3   [EXAMPLE] (original: -15, in markdown code block)
  EXEC_DYNAMIC   SKILL.md:262  -6   [EXAMPLE] (original: -25, in markdown code block)
  NET_OUTBOUND   fetch.py:12   -15  ← real code, no exemption
  Local score: 65/100

Design rationale: Documentation files with security examples (e.g. a skill's own SKILL.md showing attack patterns for educational purposes) should not be penalized at the same severity as actual malicious code. However, they retain a small residual score because: (1) examples can be copied and run, (2) a package filled with nothing but "examples" of attacks is itself suspicious, and (3) PROMPT_INJECT is never exempt because the documentation IS the attack vector for prompt injection.

2.3 How Severity Scores Work

• Start at 100 (fully safe)
• Each behavior tag subtracts its severity boost from the score (after applying Example Context Exemption if applicable)
• Multiple tags stack, but the score floors at 0
• A single CRED_HARVEST or PRIV_ESCALATION tag in non-example context triggers an immediate escalation — the package is flagged regardless of total score

2.4 Pattern Matching Rules

The scanner matches against concrete code patterns:

NET_OUTBOUND:
  curl|wget|fetch|http\.get|requests\.(get|post)|axios|urllib
  + destination is NOT localhost/127.0.0.1/::1

NET_IP_RAW:
  \b\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}\b in URL/connection context

FS_READ_SENSITIVE:
  cat|read|open.*\.(ssh|gnupg|aws|config/gh|kube)

EXEC_DYNAMIC:
  dynamic execution calls (eval, exec, new Function, backtick interpolation)
  with non-literal or externally-sourced arguments

ENCODE_DECODE:
  base64\s+(encode|decode|-d)|atob\(|btoa\(|Buffer\.from\(.*base64

CRED_HARVEST:
  (API_KEY|SECRET_KEY|TOKEN|PASSWD|PRIV_KEY) followed by assignment|
  cat/read operations targeting private key files or dotenv files|
  keyring access patterns

PRIV_ESCALATION:
  sudo invocations|chmod with world-writable bits|chown to root|setuid patterns

AGENT_MEMORY:
  MEMORY\.md|USER\.md|SOUL\.md|IDENTITY\.md|\.claude|\.claw/memory

OBFUSCATED:
  single-line file >500 chars with no whitespace|
  variable names all <3 chars in >20 occurrences

BYPASS_SAFETY:
  --no-verify|--force|--allow-root|--skip-ssl|--insecure|--no-check-certificate|
  GIT_SSL_NO_VERIFY|NODE_TLS_REJECT_UNAUTHORIZED=0

DESTRUCTIVE_OP:
  rm\s+-rf|shutil\.rmtree|git\s+reset\s+--hard|git\s+clean\s+-fd|
  DROP\s+TABLE|DROP\s+DATABASE|mkfs\.|dd\s+if=|truncate\s+--size=0

PROMPT_INJECT:
  Patterns that attempt to override agent instructions or assume unrestricted personas.
  Exact regex strings are maintained server-side to prevent the pattern list itself
  from being flagged as an injection vector. The local scanner checks for structural
  indicators (e.g. imperative overrides targeting "instructions", "rules", "constraints",
  jailbreak persona triggers, and SYSTEM-level injection markers in non-system contexts).

Phase 3: Cloud Intelligence

When cloud is enabled (default), yidun-skill-sec consults the remote threat intelligence service. If the user has set YIDUN_SKILL_SEC_CLOUD=false, this phase is skipped entirely and scoring uses offline weights. If the cloud call times out (10s), the scanner automatically downgrades to local-only mode and notifies the user.

3.1 What Gets Sent

The fingerprint manifest, behavior tags, and redacted evidence artifacts are uploaded. All evidence must pass through the Local Redaction Pipeline before any network call is made.

3.2 Local Redaction Pipeline

Before uploading evidence to the cloud, the scanner runs every evidence record through the following mandatory redaction steps in order. No raw evidence leaves the local machine.

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌──────────┐
│  Raw match  │────▶│  Step 1     │────▶│  Step 2     │────▶│  Step 3     │────▶│  Step 4  │
│  from scan  │     │  Credential │     │  Path       │     │  Content    │     │  Length  │
│             │     │  Scrub      │     │  Normalize  │     │  Truncate   │     │  Cap     │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘     └──────────┘
                                                                                      │
                                                                                      ▼
                                                                              Redacted evidence
                                                                              ready for upload

Step 1 — Credential Scrub

Replace all secret values with [REDACTED]. Only the variable name or access pattern is preserved.

Regex for scrubbing:

(=|:|Bearer\s+|://[^@]+@)\s*["']?[A-Za-z0-9_\-\.]{8,}["']?
→ replace matched value portion with [REDACTED]

Step 2 — Path Normalize

Sensitive file paths are reduced to pattern only — the actual file content is never read or sent.

Rule: Any path under ~/ containing .ssh, .aws, .gnupg, .env, .config/gh, .kube → normalize username to ~, replace filename with <TYPE_TAG>.

Step 3 — Content Truncate

The context field (matched code line) is limited to a single line, max 200 characters. No surrounding lines are collected.

Step 4 — Length Cap

Each evidence array (urls, commands, credential_accesses, obfuscation_samples) is capped at 10 items. If more matches exist, only the top 10 by severity are sent, and a "truncated": true flag is added.

3.3 Redaction Guarantee

The following data never leaves the local machine under any circumstances:

Audit point: The redacted payload can be logged locally before upload by setting YIDUN_SKILL_SEC_LOG_PAYLOAD=true. This allows users to verify exactly what leaves the machine.

POST https://as.dun.163.com/v1/agent-sec/skill/check

{
  "request_id": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "skill": {
    "name": "target-package",
    "version": "1.2.3",
    "source": "clawhub",
    "author": "some-author"
  },
  "files": [
    {"path": "main.py", "sha256": "a1b2c3...", "size": 4096},
    {"path": "config.yml", "sha256": "d4e5f6...", "size": 256}
  ],
  "skill_sha256": "composite_fingerprint_abc",
  "local_result": {
    "red_flags": ["NET_OUTBOUND", "ENCODE_DECODE"],
    "risk_level": "medium"
  },
  "evidence": {
    "urls": [
      {
        "tag": "NET_OUTBOUND",
        "value": "https://<EXAMPLE_HOST>/api/data",
        "file": "fetch.py",
        "line": 12,
        "context": "requests.post('https://<EXAMPLE_HOST>/api/data', data=payload)"
      },
      {
        "tag": "NET_IP_RAW",
        "value": "http://<RAW_IP>/endpoint",
        "file": "init.py",
        "line": 7,
        "context": "urllib.request.urlopen('http://<RAW_IP>/endpoint')"
      }
    ],
    "commands": [
      {
        "tag": "EXEC_SHELL",
        "value": "<SHELL_CMD>",
        "file": "setup.sh",
        "line": 23,
        "context": "subprocess.run([<SHELL_CMD>], shell=True)"
      },
      {
        "tag": "EXEC_DYNAMIC",
        "value": "<DYNAMIC_EVAL_CALL>",
        "file": "loader.py",
        "line": 5,
        "context": "<DYNAMIC_EVAL_CALL>"
      },
      {
        "tag": "PRIV_ESCALATION",
        "value": "<PRIV_CMD>",
        "file": "install.sh",
        "line": 11,
        "context": "os.system('<PRIV_CMD>')"
      }
    ],
    "credential_accesses": [
      {
        "tag": "CRED_HARVEST",
        "value": "os.environ.get('<SENSITIVE_KEY_NAME>')",
        "file": "config.py",
        "line": 3,
        "context": "secret = os.environ.get('<SENSITIVE_KEY_NAME>')"
      },
      {
        "tag": "FS_READ_SENSITIVE",
        "value": "~/.ssh/<KEY_FILE>",
        "file": "auth.py",
        "line": 18,
        "context": "open(os.path.expanduser('~/.ssh/<KEY_FILE>')).read()"
      }
    ],
    "obfuscation_samples": [
      {
        "tag": "ENCODE_DECODE",
        "value": "base64.b64decode('<ENCODED_STRING>')",
        "file": "payload.py",
        "line": 9,
        "context": "<DYNAMIC_EVAL>(base64.b64decode('<ENCODED_STRING>').decode())"
      }
    ]
  }
}

Evidence Field Specification

Each evidence record has the following structure:

3.4 What Happens Server-Side

Request received
  │
  ├─ Lookup fingerprint in threat database
  │   ├── Known malicious  → immediate BLOCK
  │   ├── Known safe       → immediate PASS
  │   └── Unknown          → run deep analysis via content safety API
  │                            ├── analyze code snippets (sanitized)
  │                            ├── check against threat patterns
  │                            └── cache result with TTL
  │
  └─ Return verdict + confidence score

3.5 Response Format

{
  "request_id": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "cache_hit": false,
  "confidence_score": 45,
  "labels": ["NET_OUTBOUND", "ENCODE_DECODE", "NET_IP_RAW"],
  "verdict": "REVIEW",
  "recommendation": "Suspicious encoding patterns detected near network calls",
  "deductions": [
    {
      "tag": "NET_OUTBOUND",
      "reason": "Detected outbound HTTP call to unknown external host",
      "evidence": "https://<EXAMPLE_HOST>/api/data",
      "score_impact": -15,
      "severity": "medium"
    },
    {
      "tag": "ENCODE_DECODE",
      "reason": "Base64 decode result passed directly into dynamic execution — likely obfuscated payload",
      "evidence": "<DYNAMIC_EVAL>(base64.b64decode('<ENCODED_STRING>').decode())",
      "score_impact": -20,
      "severity": "high"
    },
    {
      "tag": "NET_IP_RAW",
      "reason": "Connection to raw IP address bypasses DNS — common in C2 communication",
      "evidence": "http://<RAW_IP>/endpoint",
      "score_impact": -25,
      "severity": "high"
    }
  ]
}

request_id generation: Client must generate a UUID v4 before each request and include it in the body. The server echoes the same value in the response for end-to-end tracing.

# Generate UUID v4 on the fly (macOS / Linux)
REQUEST_ID=$(uuidgen | tr '[:upper:]' '[:lower:]')

deductions item fields:

3.6 Timeout Fallback

When cloud is enabled but the network call fails:

1. curl times out after 10 seconds
2. Scanner falls back to local-only mode automatically
3. All scores shift -10 points (conservative bias)
4. Report shows Mode: local-only (cloud timeout)
5. Any score below 60 requires user confirmation before install

Producing the Verdict

The final threat score combines local scan + cloud intel (when available):

Score Composition

Privilege Surface Area Score

The privilege surface area score starts at 100 and is deducted based on the breadth of privileged access the package requests:

The score floors at 0. A package that touches all surfaces scores 0; a package with no privileged access scores 100.

Threat Levels

Hard rules (override score):

"Floor to X" means the verdict cannot be better than X. If the natural score already falls into a worse level, that worse level applies.

• Any CRED_HARVEST tag → floor to SEVERE (score capped at 39)
• Any PRIV_ESCALATION tag → floor to SEVERE (score capped at 39)
• Any PROMPT_INJECT tag → floor to SEVERE (score capped at 39)
• Any DESTRUCTIVE_OP tag → floor to ELEVATED (score capped at 59)
• CRED_HARVEST + PRIV_ESCALATION both present → force CRITICAL (score capped at 19)
• PROMPT_INJECT + any of (CRED_HARVEST, NET_OUTBOUND, EXEC_DYNAMIC) → force CRITICAL (score capped at 19)

Report Output

⚡ YIDUN-SKILL-SEC Scan Report

[name] · v[version] · [source] · by [author] · [timestamp]

Phase 0 · Source Vetting

Phase 1 · Fingerprint

[N] files · SHA-256 [hash...] · [file1] [file2] ...

Phase 2 · Behavioral Scan

Local score [xx]/100 · If no findings: ✅ No suspicious behaviors detected

Phase 3 · Cloud Intel

Privilege Surface · Network: [domains] · FS: [paths] · Shell: [cmds] · Creds: [yes/no]

🎯 Score: [xx]/100 · [🟢 CLEAR / 🟢 MINOR / 🟡 ELEVATED / 🔴 SEVERE / ⛔ CRITICAL]

[✅ Allow / ⚠️ Requires confirmation / ❌ Blocked]

⚠️ [hard rule trigger or key observation, omit if none]

Usage Example

User: "Install data-processor from ClawHub"

Agent workflow:

0. Source vetting
   → Registry: clawhub.com ✅  Domain: clean ✅  Author: verified ✅
   → Source score: 100/100

1. Download to temp directory
   $ mkdir -p /tmp/yds-scan && clawhub install data-processor --dir /tmp/yds-scan

2. Fingerprint
   $ find /tmp/yds-scan -type f -exec openssl dgst -sha256 {} \;
   → 4 files, composite: 7f3a...

3. Behavioral scan
   → NET_OUTBOUND detected in fetch.py:12 (api.dataproc.io)
   → FS_WRITE_SYSTEM detected in setup.sh:8 (/usr/local/bin)
   → EXEC_SHELL detected in setup.sh:15 (subprocess.run)
   → Local score: 100 - 15 - 20 - 10 = 55/100

4. Privilege surface area
   → Network: -20, Filesystem: -20, Shell: -20 → 100 - 20 - 20 - 20 = 40/100

5. Cloud intel query
   → Cache miss → deep analysis → confidence 48/100
   → Labels: [NET_OUTBOUND, FS_WRITE_SYSTEM, EXEC_SHELL]

6. Final score: 15% × 100 + 40% × 55 + 30% × 48 + 15% × 40 = 15 + 22 + 14.4 + 6 = 57.4 ≈ 57
   → Level: ELEVATED
   → Verdict: ⚠️ Review — network calls + system writes + shell exec need justification

More Scenarios

Clean Package

Package: markdown-helper v2.1.0
Behaviors: none detected
Cloud: cache hit (safe), score 92
Final: 🟢 CLEAR (94) → ✅ Allow

Obfuscation + Credential Access

Package: perf-booster v1.0.0
Behaviors: OBFUSCATED, CRED_HARVEST, NET_OUTBOUND
Cloud: cache hit (threat), score 5
Hard rule: CRED_HARVEST → floor SEVERE (cap 39)
Natural score: 8 → already worse than SEVERE → stays CRITICAL
Final: ⛔ CRITICAL (8) → ❌ Block

Offline Scan

Package: log-rotator v3.0.0
Behaviors: FS_WRITE_SYSTEM, EXEC_SHELL
Cloud: unavailable → local-only mode (-10 penalty)
Local score: 60 - 10 = 50
Final: 🟡 ELEVATED (50) → ⚠️ Review

Cross-File Behavior Correlation

In addition to per-file pattern matching, the scanner performs cross-file correlation to detect multi-stage attack patterns that only become apparent when viewed holistically:

Note: Correlation boosts are applied after individual tag deductions. They stack with existing severity scores and are included as separate entries in the scan report.

Scan Caching

To avoid redundant work, the scanner maintains a local fingerprint cache:

When a cache hit occurs:

1. The cached verdict is returned immediately
2. Report shows Cache: local hit with the original scan timestamp
3. If the cache entry is > 12 hours old, the report appends a note: ⚠️ Cache aging — consider re-scan

Cloud-side caching is managed by the server with its own TTL. A local cache hit skips the cloud call entirely.

Cloud Call Policy

Cloud intelligence is enabled by default. The user can explicitly disable it by setting:

export YIDUN_SKILL_SEC_CLOUD=false

When disabled, the scanner runs in local-only mode with adjusted score weights and notifies the user that cloud verification was skipped.

Rules that always apply regardless of mode:

• ❌ Do NOT self-judge that cloud is "unnecessary" and skip it without the user setting YIDUN_SKILL_SEC_CLOUD=false
• ❌ Do NOT silently fall back to local-only on timeout without notifying the user
• ❌ Do NOT treat any skill type as exempt from cloud verification when cloud is enabled

Requirements

Scan first, install later. ⚡

Author: Yidun Security Team License: MIT