Back to skill

Security audit

Principles Agent

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed first-principles task-planning helper with no direct network access, dependency install, credential access, or persistence, though its validation flow appears unreliable and should not be treated as a safety gate.

Install this only if you want a Chinese-language planning/orchestration skill that may make multiple LLM calls through OpenClaw's main agent. Treat its quality validation as advisory because the inspected code appears to parse prompt text in some validation paths rather than a separate validator response. Use the --output option only with paths where you intend to write a report.

Vulnerability Patterns
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (25)

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

validate_task is documented as performing validation, but it never invokes a model or validator on the constructed prompt. Instead it passes the prompt text directly into _parse_response, which attempts to extract JSON from arbitrary task/output content; this can cause validation to fail unpredictably, crash on json.loads, or misinterpret embedded JSON in untrusted output as an authoritative validation result.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The inline comment states that the actual LLM call occurs elsewhere, but the function immediately returns _parse_response(prompt). This creates a logic flaw where security or quality gating can be bypassed by implementation error: callers may believe output was validated when no validation actually occurred, undermining any workflow that depends on validation_passed before proceeding.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill documentation, usage examples, and user-facing text are entirely in Chinese, and the README does not indicate that other languages are supported or that Chinese is an optional locale. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
71% confidence
Finding

文件标题及全文说明均以中文呈现,并将技能定位为中文描述的使用方式,但未声明支持其他语言或允许用户选择输出语言。若组织要求避免未经用户选择而固定语言/locale,这种默认单一语言行为可能构成自然语言策略违规。

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger conditions are broad enough to overlap with many normal planning, analysis, or design requests, which can cause the skill to activate unexpectedly. In an agent framework, over-broad activation can override user intent, route sensitive prompts through unnecessary decomposition logic, and increase the chance of unsafe downstream tool use by the main agent even if this skill itself does not directly perform I/O.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The top-level docstring presents the skill name, purpose, and usage entirely in Chinese, which imposes a specific language on users without any opt-in or alternative locale. This is a natural-language policy concern because the file does not indicate that the skill is region-specific or provide a user-selectable language option.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The CLI description and argument help text are all written in Chinese, which forces a single language for user-facing interaction. There is no indication that this is a region-specific tool or that users can opt into another language, so it falls under the language/locale policy violation category.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

This Python file contains user-facing natural-language documentation entirely in Chinese, including the title and function docstring, with no indication that the user can choose another language or that the skill is region-specific. Under the policy, forcing a specific language or locale without opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest’s user-facing permission descriptions and responsibility summary are written entirely in Chinese, while the rest of the package metadata is in English. This creates a language/locale constraint without any opt-in or explanation that the skill is intended only for Chinese-speaking users or a China-specific environment.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The constructed prompt is entirely in Chinese and instructs the downstream model in that language, which imposes a specific language/locale behavior on task execution. There is no visible opt-in, fallback, or justification that this executor is intended only for Chinese-language use.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The method docstring at L103 says it will '并行执行一批无依赖的任务' (execute a batch of independent tasks in parallel), but the implementation uses a normal for-loop and calls the synchronous executor one task at a time. This is an active contradiction between the documented intent and actual behavior, not merely an omitted detail.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The embedded system prompt is entirely in Chinese and directs the model to produce the final integrated result in that language/format, but the file does not offer any user language choice or explain a required locale constraint. This is a natural-language locale policy issue because it imposes a specific language behavior by default.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The orchestrator automatically sends generated subtasks to an external executor for parallel execution with no consent gate, capability restriction, or visible disclosure at the orchestration layer. Because the subtasks are derived from LLM output and prior prompt content, this creates a dangerous path from untrusted input to real-world actions, increasing the risk of unintended external operations, data exposure, or tool misuse.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The module docstrings and embedded system prompt are written entirely in Chinese and direct the model's behavior in that language, but there is no indication that language selection is optional or limited to a China-specific use case. This creates a natural-language locale policy issue because the skill effectively imposes a specific language without user opt-in.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The module docstrings and the SYSTEM_PROMPT require operation in Chinese only, and the file does not offer any mechanism for the user to select another language or locale. This is a natural-language policy concern because it imposes a specific language without documented opt-in or justification.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The method decompose is documented and commented as part of an LLM-based task decomposition flow, with line L059 explicitly stating the actual LLM call occurs in an upper-layer orchestrator. However, the code immediately passes the generated prompt string into _parse_response, which expects JSON and is described as parsing an LLM response. This is an active contradiction between the documented intent and implemented behavior.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The docstring at L068 says the method re-decomposes tasks according to feedback, implying an LLM-assisted refinement step. In practice, it constructs a prompt with feedback and then calls _parse_response(prompt) without any model invocation, so it does not perform the documented decomposition behavior. This directly conflicts with the stated intent of the method.

Content

No source excerpt is available for this finding.

Ssd 4

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

User-controlled goal, truths, and feedback are interpolated directly into the same prompt channel as trusted system guidance, allowing prompt injection and instruction smuggling. An attacker can embed adversarial directives in these fields to steer decomposition output, override intended constraints, or cause downstream agents to generate unsafe or policy-bypassing subtasks, especially if later components trust those subtasks.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file’s natural-language instructions, including the system prompt and surrounding descriptions, are entirely in Chinese and implicitly require the model interaction to occur in that language. There is no indication that the user can choose a different language or that the locale restriction is justified as region-specific, which matches the language/locale policy violation criteria.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The method is documented as parsing an LLM response, but it actually passes the locally constructed prompt into _parse_response. This breaks the trust boundary and can cause user-controlled goal content to be interpreted as structured JSON if it contains braces or JSON-like text, leading to incorrect outputs, parser failures, or prompt-injection-style data confusion in downstream orchestration.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The re-derivation path repeats the same flaw: it builds a prompt incorporating attacker-influenced feedback and then parses that prompt as if it were LLM output. This allows untrusted input in feedback or goal fields to directly influence the JSON parser path, potentially causing denial of service, malformed truth objects, or corrupted reasoning state.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The system prompt is entirely in Chinese and mandates the model's validation behavior and output format in that language. This is a natural-language locale policy concern because the file does not offer a language choice or explain that the skill is intentionally region-specific.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The only natural-language description in this file is written entirely in Chinese, with no indication that users can choose another language or that the skill is intentionally limited to a Chinese-speaking context. Under the policy, forcing a specific language without user opt-in can be a locale/language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

The file-level docstring and class/method descriptions are written entirely in Chinese, indicating a fixed language choice in the skill's natural-language interface. Under the policy, forcing a specific language without user opt-in or a documented regional justification is a locale-policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The top-level docstring describes the orchestrator exclusively in Chinese, and the rest of the file continues this pattern for user-facing prompts and reports. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation unless it is explicitly justified as region-specific.

Content

No source excerpt is available for this finding.