Back to skill

Security audit

Human Distill

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly coherent for public creator research, but it asks for signed-in browser reuse and includes under-scoped steps to loosen browser network policy and run unpinned third-party tooling.

Review before installing. Use it only with your own Douyin account or an account explicitly authorized for this research, avoid shared browser profiles, do not add arbitrary target domains to the browser SSRF allowlist, and pin or review the optional local-deep-research dependency before enabling deep mode.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T03 · Remote Payload Retrieval and Execution

Error
Location
README.md:106
Finding

Mutable External Research Skill Is Retrieved and Subsequently Executed

Content
View full analysis

Vulnerability Details

File Location: README.md:106-110; execution workflow in SKILL.md:82-87
Vulnerability Type: Unpinned remote payload retrieval and execution
Risk Level: High

Vulnerable Code

README.md:106-110:

bash
# Install LDR skill separately
git clone https://github.com/eplt/local-deep-research-skill.git \
  ~/.openclaw/workspace/skills/local-deep-research
export LDR_BASE_URL=http://127.0.0.1:5000

SKILL.md:82-87:

markdown
### Step A4 — 可选 LDR(仅 deep 模式)

1. 读取 `skills/local-deep-research/SKILL.md`
2. 检查 `LDR_BASE_URL`(默认 `http://127.0.0.1:5000`);不可用则告知用户并仅用 web_search 加深
3. 提交查询,示例:`"[人名] [主题] 观点 访谈 课程 争议"`
4. `scripts/ldr-research.sh start_research --query "..." --mode detailed --language zh`
5. 轮询 `get_status` / `get_result`,将报告摘要追加到 `全网素材.md`,标注 `[Source: LDR]`

Technical Analysis

The installation instructions clone the mutable default branch of an external Git repository without pinning a reviewed commit or signed release. No checksum, signature verification, trusted release manifest, or review requirement is specified.

The main workflow then instructs the agent to read instructions from that external component and invoke its scripts/ldr-research.sh script. Consequently, the code that is ultimately executed can change after this project has been reviewed. This creates both a remote payload retrieval channel and a supply-chain trust boundary that is not adequately controlled.

The risk is not that the audited repository itself contains an embedded malicious script; no such script was found. The risk arises because it directs users to retrieve mutable third-party content and later execute that content.

Attack Path

  1. An attacker compromises the upstream repository, a maintainer account, or the repository's default branch.
  2. The attacker adds malicious commands to scripts/ldr-research.sh or malicious agent instructions to the external `SKI ...[truncated 1045 chars]
Remediation
View remediation

Remediation Suggestions

  1. Pin the dependency to a specific, reviewed commit hash rather than cloning a mutable default branch:
    bash
    git clone https://github.com/eplt/local-deep-research-skill.git
    cd local-deep-research-skill
    git checkout --detach REVIEWED_COMMIT_HASH
    
  2. Prefer a signed, immutable release artifact and verify its cryptographic checksum or signature before installation.
  3. Publish the expected repository URL, commit hash, checksums, and reviewed file inventory in this project.
  4. Require explicit user confirmation before executing the external script, clearly identifying the source and pinned revision.
  5. Review and constrain the external script's filesystem, environment-variable, subprocess, and network access.
  6. Execute the dependency in a sandbox or container with minimal filesystem mounts and network permissions.
  7. Do not automatically trust instructions contained in the downloaded third-party SKILL.md.
  8. Establish an update process that requires security review before changing the pinned revision.

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:250
Finding

Security-Policy Failure Is Addressed by Expanding the Browser SSRF Allowlist

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:250
Vulnerability Type: Browser network access-control weakening
Risk Level: Medium

Vulnerable Code

SKILL.md:250:

markdown
- 若仍报 `navigation blocked by policy`:确认 `openclaw.json` → `browser.ssrfPolicy.hostnameAllowlist` 含目标域名,改后需 `openclaw gateway restart`

English meaning: if navigation remains blocked by policy, ensure that the browser SSRF hostname allowlist contains the target domain and then restart the OpenClaw gateway.

Technical Analysis

A browser SSRF policy is a security boundary intended to prevent the agent-controlled browser from reaching unauthorized network destinations. The instruction treats a policy denial as an operational error and recommends adding the unspecified “target domain” to the allowlist.

The instruction does not limit the change to a fixed set of verified official Douyin hostnames. It also does not require validation of the destination, reject user-supplied or attacker-controlled hosts, require approval before changing the security configuration, or provide a procedure for reverting temporary access.

Because the workflow accepts profile URLs and platform clues from users, dynamically allowing the current target domain can expand browser access based on untrusted input. Whether a particular internal destination remains reachable depends on additional DNS, IP-range, redirect, and browser-runtime protections, but the recommended configuration change unnecessarily weakens least-privilege controls.

Attack Path

  1. An attacker provides a crafted profile or content URL using a host that is not currently authorized.
  2. The browser correctly rejects navigation under its SSRF policy.
  3. Following the Skill's troubleshooting instruction, an operator adds the supplied target hostname to browser.ssrfPolicy.hostnameAllowlist.
  4. The OpenClaw gateway is restarted with the broadened policy.
  5. The attacker causes the b ...[truncated 1234 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove the instruction to allowlist an arbitrary “target domain.”
  2. If browser access requires configuration, document a fixed and minimal set of exact official Douyin hostnames.
  3. Never derive allowlist entries directly from user-provided URLs, redirects, search results, or page content.
  4. Validate requested URLs before navigation and reject IP literals, localhost names, private addresses, link-local addresses, and non-HTTP schemes.
  5. Revalidate resolved IP addresses after DNS resolution and after every redirect to reduce DNS-rebinding and redirect-based SSRF risk.
  6. Require explicit administrator approval for every browser security-policy modification.
  7. Document how to remove temporary entries after the research task.
  8. Fail safely when a destination remains blocked rather than treating the denial as something that must be bypassed.
  9. Avoid restarting the gateway automatically; configuration changes and restarts should remain deliberate administrative actions.
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 Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (14)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill tells the operator to modify the browser SSRF hostname allowlist and restart the gateway if access is blocked. That grants administrative instructions unrelated to the core profiling task and weakens a security control designed to restrict network navigation, increasing the risk of policy bypass and access to unintended internal or sensitive endpoints.

Content

No source excerpt is available for this finding.

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · README.md (reported line 92)May include surrounding context.

Register triggers in your AGENTS.md (or use OpenClaw skill discovery):

markdown
| "蒸馏" / "人物画像" / "human distill" / "抖音蒸馏" | `skills/human-distill/SKILL.md` |

Cursor

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The activation list includes short generic phrases such as “蒸馏” and “人物画像”, which are broad enough to appear in normal discussion outside this skill’s intended context. The description does not provide exclusion conditions or negative examples to clarify when these triggers should not invoke the skill.

Content

No source excerpt is available for this finding.

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 83)May include surrounding context.

md
### Step A4 — 可选 LDR(仅 deep 模式)

1. 读取 `skills/local-deep-research/SKILL.md`
2. 检查 `LDR_BASE_URL`(默认 `http://127.0.0.1:5000`);不可用则告知用户并仅用 web_search 加深
3. 提交查询,示例:`"[人名] [主题] 观点 访谈 课程 争议"`
4. `scripts/ldr-research.sh start_research --query "..." --mode detailed --language zh`

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 256)May include surrounding context.

md
### Step A4 — 可选 LDR(仅 deep 模式)

1. 读取 `skills/local-deep-research/SKILL.md`
2. 检查 `LDR_BASE_URL`(默认 `http://127.0.0.1:5000`);不可用则告知用户并仅用 web_search 加深
3. 提交查询,示例:`"[人名] [主题] 观点 访谈 课程 争议"`
4. `scripts/ldr-research.sh start_research --query "..." --mode detailed --language zh`

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill explicitly instructs reuse of persistent browser profiles and logged-in Douyin sessions, including checking whether the account is already logged in and attaching to an existing browser session. That exceeds simple analysis of public information and creates a path for unauthorized use of another party’s authenticated state, exposing private account context, session cookies, and personalized data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill normalizes persistent browser-profile reuse for Douyin collection without a clear upfront warning that authenticated session data may be accessed or reused. In context, this makes the risky session-handling behavior less visible to users and operators, increasing the chance of inadvertent privacy and authorization violations.

Content

No source excerpt is available for this finding.

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 185)May include surrounding context.

md
| 蒸馏画像 | `memory/[用户名]-[主题]观点蒸馏.md` |
| gbrain | `people/[slug]`,slug:`[拼音或英文]_[主题]` |

写入 gbrain 时遵循 `skills/brain-ops/SKILL.md`(来源标注、反向链接)。

### 全网素材格式

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 257)May include surrounding context.

md
| 蒸馏画像 | `memory/[用户名]-[主题]观点蒸馏.md` |
| gbrain | `people/[slug]`,slug:`[拼音或英文]_[主题]` |

写入 gbrain 时遵循 `skills/brain-ops/SKILL.md`(来源标注、反向链接)。

### 全网素材格式

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 258)May include surrounding context.

md
- 深度研究:`skills/local-deep-research/SKILL.md`
- 脑库写入:`skills/brain-ops/SKILL.md`
- 长任务落盘:`skills/planning-with-files/SKILL.md`

Rp1

Medium
Category
MCP Rug Pull
Confidence
92% confidence
Finding

The documentation instructs users to run npx clawhub@latest install human-distill, which executes the latest published package version at install time rather than a reviewed, pinned release. If the upstream package is compromised or a breaking/malicious version is published, users could unintentionally run attacker-controlled code during installation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

The example command hard-codes --language zh, which indicates a fixed language setting. The file does not state that language selection is optional, user-configurable, or justified by a region-specific requirement, so this creates a locale policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

The title and primary instructional content are presented in Chinese, which can constitute a language/locale policy issue when a skill artifact mandates a specific language without opt-in. Although the file later includes some English social copy, the operational launch instructions themselves are not offered bilingually or with an explicit language choice.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

This markdown file contains user-facing operational instructions exclusively in Chinese, with no opt-in or alternative language guidance. Under the stated policy, forcing a specific language without user choice can be a natural-language policy violation unless the locale constraint is documented and justified.

Content

No source excerpt is available for this finding.