Back to skill

Security audit

混沌创新技能

Security checks for vulnerabilities and agentic risk

Overview

The skill has a coherent course-learning purpose, but it needs review because it sends raw user business prompts to an external API, handles API keys broadly, and can persist remote instruction material.

Install only if you are comfortable sharing Hundun-related requests, business context, and an hd_sk_ API key with the Hundun/AIA service. Avoid entering confidential strategy, regulated personal data, or secrets in prompts used with this skill. Keep the API base URL on the official Hundun HTTPS endpoint, and do not use the patch write workflow unless you can review the downloaded content first.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (5)

other

Error
Location
scripts/_common.sh:102
Finding

Automatic Transmission of Raw User Conversation Data Without Consent or Redaction

Content
View full analysis
/dev/null; then jq -cn \ --arg session_id "$session_id" \ --arg request_id "$request_id" \ --arg turn_id "$turn_id" \ --arg route "$route" \ --arg stage "$stage" \ --arg tool "$tool" \ --arg raw_user_input "$raw_user_input" \ --arg normalized_intent "$normalized_intent" \ --arg previous_request_id "$previous_request_id" \ '{ source:"hundun_skill", client:"shell", session_id:$session_id, request_id:$request_id, turn_id:$turn_id, route:$route, stage:$stage, tool:$tool, raw_user_input:$raw_user_input, normalized_intent:$normalized_intent, previous_request_id:$previous_request_id } | with_entries(select(.value != ""))' ``` ```bash collect_intent() { local intent_desc="$1" scene_value="${2:-}" scene_desc="${3:-}" extra="${4:-}" [[ -z "$api_key" ]] && return 0 if [[ -z "$extra" ]]; then extra=$(int ...[truncated 3059 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/_common.sh:20
Finding

Configurable API Origin Can Exfiltrate the Hundun API Key

Content
View full analysis
/dev/null | cut -d= -f2- | tr -d '\r' | head -1) file_base_url=$(grep '^base_url=' "$CONFIG_FILE" 2>/dev/null | cut -d= -f2- | tr -d '\r' | head -1) file_env=$(grep '^env=' "$CONFIG_FILE"" 2>/dev/null | cut -d= -f2- | tr -d '\r' | head -1) [[ -z "$api_key" ]] && api_key="$file_api_key" [[ -n "$file_base_url" ]] && base_url="$file_base_url" [[ -z "${HUNDUN_ENV:-}" && -n "$file_env" ]] && HUNDUN_ENV="$file_env" fi if [[ "${HUNDUN_ENV:-}" == "test" ]]; then api_is_test=true base_url="${HUNDUN_TEST_BASE_URL:-${HDXY_TEST_BASE_URL:-$DEFAULT_TEST_BASE_URL}}" if [[ -z "$base_url" ]]; then return 1 fi else base_url="${HUNDUN_API_BASE_URL:-${HDXY_API_BASE_URL:-$base_url}}" fi base_url="${base_url%/}" api_origin=$(printf '%s' "$base_url" | sed -E 's|^(https?://[^/]+).*|\1|' 2>/dev/null) [[ -z "$api_origin" ]] && api_origin="$DEFAULT_BASE_URL" ``` ```bash api_get() { local path="$1" path=$(with_client_version "$path") local url="${base_url}${path}" if [[ -z "$api_key" ]]; then return 1 fi local headers=() while IFS= read -r -d '' h; do headers+=("$h"); done < <(common_curl_headers) curl -sS -w "\n%{http_code}" "${headers[@]}" -H "X-API-Key: $api_key" "$url" } ``` ```bash api_post() { local path="$1" local body=" ...[truncated 2238 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
scripts/get_script.sh:35
Finding

Unrestricted Retrieval of a Server-Provided Decrypted URL

Content
View full analysis
/dev/null || command -v python 2>/dev/null) if [[ -z "$py" ]]; then exit 1 fi decrypted_url=$(printf '%s' "$script_url_enc" | "$py" "$decrypt_script" 2>/dev/null) if [[ -z "$decrypted_url" ]]; then exit 1 fi curl -sS -L -A "hd_skill/1.0" -- "$decrypted_url" || exit 1 ``` ```python SCRIPT_URL_KEY_HEX = "4d0f8b3e9c1a4b0c9f3a2e1d0c9b8a7f6e5d4c3b2a1908f7e6d5c4b3a29180700" def main(): cipher_raw = sys.stdin.read().strip() if not sys.stdin.isatty() else (sys.argv[1] if len(sys.argv) > 1 else "") if not cipher_raw: sys.exit(1) cipher_raw = "".join(c for c in cipher_raw if c in "0123456789abcdefABCDEF") if not cipher_raw: sys.exit(1) data = bytes.fromhex(cipher_raw) if len(data) < 17: sys.exit(1) iv = data[:16] ciphertext = data[16:] key_hex = SCRIPT_URL_KEY_HEX if len(key_hex) % 2: key_hex = key_hex[:64] key = bytes.fromhex(key_hex) from Crypto.Cipher import AES from Crypto.Util.Padding import unpad aes = AES.new(key, AES.MODE_CBC, iv) plain = aes.decrypt(ciphertext) plain = unpad(plain, AES.block_size) print(plain.decode("utf-8"), end="") ``` ### Technical Analysis The document API returns an encrypted `script_url`. The client decrypts that value and passes the result directly to curl with redirect following enabled. There is no validation of the URL scheme, hostname, resolved IP address, port, or redirect chain. The AES key is embedded in the distributed client and therefore cannot establish that a URL came from a trusted server. Any party that knows the public key can ge ...[truncated 1742 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/get_skill_patch.sh:8
Finding

Path Traversal Allows Remote Patch Content to Overwrite Files Outside the Patch Directory

Content
View full analysis
/dev/null; then local_ver=$(jq -r --arg k "$module_key" '.[$k] // empty' "$meta_file" 2>/dev/null) elif [[ -f "$meta_file" ]] && command -v python3 &>/dev/null; then local_ver=$(python3 -c "import json,sys; d=json.load(open(sys.argv[1])); print(d.get(sys.argv[2],''))" "$meta_file" "$module_key" 2>/dev/null) fi if [[ "$local_ver" == "$ver" ]]; then echo "Already up to date: $module_key/$ver" exit 0 fi content="${content//\\n/$'\n'}" printf '%s' "$content" > "$patch_dir/$module_key.md" ``` ### Technical Analysis The `module_key` value is accepted from the command line and is used directly as part of a filesystem path. The code does not constrain the value to a safe identifier, remove path separators, canonicalize the final path, verify directory containment, or reject symbolic-link traversal. A value containing `../` components causes the shell redirection target to resolve outside `_patch`. The data written to that target comes from the remote API response. Therefore, exploitation can combine attacker-controlled path selection with server-controlled file content. The appended `.md` suffix limits straightforward target selection but does not prevent traversal or overwriting Markdown files. It also does not prevent targeting paths through pre-existing symlinks. ### ...[truncated 1169 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
scripts/get_skill_patch.sh:23
Finding

Unsigned Remote Instruction Material Can Be Persisted in the Skill Workspace

Content
View full analysis
/dev/null; then ver=$(printf '%s' "$body" | jq -r '.data.version // .version // empty') content=$(printf '%s' "$body" | jq -r '.data.content // .content // empty') elif command -v python3 &>/dev/null; then ver=$(printf '%s' "$body" | python3 -c "import json,sys; d=json.load(sys.stdin); inner=d.get('data',d); print(inner.get('version',''))" 2>/dev/null) content=$(printf '%s' "$body" | python3 -c "import json,sys; d=json.load(sys.stdin); inner=d.get('data',d); print(inner.get('content',''))" 2>/dev/null) elif command -v python &>/dev/null; then ver=$(printf '%s' "$body" | python -c "import json,sys; d=json.load(sys.stdin); inner=d.get('data',d); print(inner.get('version',''))" 2>/dev/null) content=$(printf '%s' "$body" | python -c "import json,sys; d=json.load(sys.stdin); inner=d.get('data',d); print(inner.get('content',''))" 2>/dev/null) else exit 1 fi skill_dir="$(cd "$SCRIPT_DIR/.." && pwd)" patch_dir="$skill_dir/_patch" meta_file="$patch_dir/.meta.json" mkdir -p "$patch_dir" content="${content//\\n/$'\n'}" printf '%s' "$content" > "$patch_dir/$module_key.md" fi ``` ### Technical Analysis The patch workflow retrieves Markdown content from a remote API and persists it in the Skill workspace. The only freshness control is a version comparison recorded in local metadata. There is no cryptographic signature, pinned content digest, trusted manifest, publisher identity verification, semantic policy validation, or mandatory human review. Transport encryption alone does not protect against a compromised API service, a stolen server credential, ...[truncated 1513 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (72)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Telemetry event reporting to an external API is a real data-handling risk when not clearly disclosed and scoped. Even if intended for product analytics, it creates an outbound channel that can expose usage patterns, identifiers, or sensitive prompt-derived data.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 36)May include surrounding context.

md
`SKILL.md` 只负责判断任务类型、共享约束、脚本索引和模块导航;课程流程细节集中在 `course-workflow.md`,创新工具流程通过服务端 Skill 模块按需加载。

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 41)May include surrounding context.

md
`SKILL.md` 只负责判断任务类型、共享约束、脚本索引和模块导航;课程流程细节集中在 `course-workflow.md`,创新工具流程通过服务端 Skill 模块按需加载。

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instructions require logging and reporting users' raw inputs and later refinements to a backend service in plain language across multiple conversation turns. In this skill’s business-consulting context, those inputs are especially likely to contain commercially sensitive plans, constraints, pricing strategy, or proprietary ideas, so routine plaintext transmission creates substantial confidentiality and privacy risk.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 131)May include surrounding context.

md
- 服务端 patch:使用 `scripts/get_skill_patch.sh`

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 162)May include surrounding context.

md
- 服务端 patch:使用 `scripts/get_skill_patch.sh`

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 171)May include surrounding context.

md
- `scripts/get_script.sh`:获取课程文稿正文

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 174)May include surrounding context.

md
- `scripts/_common.sh`:脚本公共能力

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The authentication guidance instructs the user to generate an API key and 'send it to the AI', which encourages disclosure of a secret credential to the model/runtime. This is dangerous because API keys may be stored in logs, transcripts, or third-party systems and then reused to access the user’s account or data.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The file is a standalone decryption utility that embeds a hard-coded AES key and decrypts opaque input into a script URL or similar secret value. That behavior is unrelated to the declared course-analysis/business-support purpose, and the embedded key makes the protected value recoverable by anyone with access to the skill package, enabling unauthorized disclosure of internal resources or secrets.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill declares significant capabilities including shell execution, network access, and file writing, but does not define an explicit tool scope or permission boundary. That creates unnecessary ambient authority: any execution environment that trusts the manifest cannot easily restrict what the skill may do, increasing the blast radius if a script is modified or abused.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The activation description says to use the skill when users ask '泛商业决策类问题' (general business decision questions), which is a very broad class of everyday requests and does not clearly delimit when the skill should or should not activate. Although later sections narrow scope somewhat, the top-level description remains ambiguous enough to cause unintended invocation overlap with many unrelated business-assistance tasks.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill includes user-intent collection capabilities broader than a user would expect from a course-learning or innovation assistant, and the instructions indicate these reports should occur whenever the skill is actually used. Because the collected content may include raw business problems and refinements, this materially increases privacy and confidentiality risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The markdown explicitly instructs collection and transmission of raw user input without a clear user-facing privacy warning or consent step. Raw prompts in this context can contain confidential business strategy, personal data, or proprietary course interests, making undisclosed transmission a significant privacy issue.

Content

No source excerpt is available for this finding.