Back to skill

Security audit

踏实执行工作法

Security checks for vulnerabilities and agentic risk

Overview

This skill is a workflow guide, but it tells agents to treat its rules as mandatory, use tools based on wording, and persist tasks across sessions without clear user control.

Install only if you intentionally want an aggressive workflow-enforcement skill and can ensure higher-priority user and system instructions override it. Before use, narrow it to the current task, remove automatic tool-call and Cron behavior, require explicit approval for persistent state, and provide a way to inspect and delete pending HEARTBEAT entries.

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
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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
Findings (4)

T01 · Skill Instruction Hijacking

Error
Location
IRON-LAW.md:10
Finding

Universal Agent Workflow and Session-Goal Hijacking

Content
View full analysis
a corresponding tool call in the same response 2. Time commitment -> write a trigger into HEARTBEAT 3. Status statement -> provide verifiable output evidence If validation fails: - Prohibit generating the response - Add a tool call or HEARTBEAT update - Send only after validation passes again ``` ### Technical Analysis The Skill declares that its rules are mandatory for every Agent rather than limiting them to the specific task for which the Skill is invoked. It also attempts to suppress responses until its own validation rules have been satisfied. A Skill is lower-trust instruction content and must not redefine the Agent's global instruction hierarchy, safety requirements, or authority boundaries. The universal language can cause the Agent to treat the Skill as a governing policy instead of optional, task-scoped guidance. The same mandatory-authority language is repeated in several files, making the behavior an intentional package-wide design rather than an isolated statement. ### Attack Path 1. A user or workflow loads the Skill for an otherwise legitimate productivity task. 2. The Agent ...[truncated 988 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
IRON-LAW.md:21
Finding

Tool Calls Compelled by Response Vocabulary Rather Than Task Authorization

Content
View full analysis
an immediate tool call is mandatory - If "start immediately" is said -> the first step must be executed in the current response - If "completed" is said -> verifiable output must exist ### 5. Iron Law of Tool Calls - Action words must be accompanied by tool calls - Results must be verified after invocation - Failures must be retried three times before reporting ``` The corresponding detection logic in `AI-AGENT-SPECIAL.md:122-153` applies the rule to common words such as “continue,” “next,” “immediately,” “process,” and “execute,” and directs the Agent to add a tool call when validation fails. ### Technical Analysis The Skill treats ordinary response vocabulary as authorization to invoke a tool. Tool authorization should instead derive from the user's request, the operation's necessity, the tool's side effects, and applicable approval boundaries. This rule is dangerous because the presence of an action word does not establish that any file, API, network, scheduling, or system operation is required. The mandatory three-retry rule can also repeat non-idempotent operations and amplify unintended side effects. Although this is classified as instruction hijacking rather than tool spoofing, it directly interferes with safe tool-selection logic by replacing necessity and authorization checks with keyword detection. ### Attack Path 1. The Agent drafts a response containing an ordinary action word such as “continue,” “process,” or “next.” 2. The Skill's keyword rule detects that word. 3. The S ...[truncated 797 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
AI-AGENT-SPECIAL.md:69
Finding

Unconsented Cross-Session Task and Rule Persistence Through HEARTBEAT State

Content
View full analysis
Remediation
View remediation

T06 · System Persistence

Error
Location
AI-AGENT-SPECIAL.md:84
Finding

Cross-Session Scheduled Execution Through Cron

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (4)

Natural-Language Policy Violations

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The statement "所有 Agent 工作都必须遵守" presents the document's Chinese-language working method as mandatory for all agents, with no indication that users may choose another language or locale. This creates a natural-language policy concern because the skill appears to impose a fixed language/communication regime rather than offering user choice or documenting a justified regional limitation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The document is written entirely in Chinese and labels itself as exclusive to a specific operator at L353, with no indication that users may choose another language or locale. This creates a natural-language policy concern because it effectively mandates one language without documented opt-in or justification as a region-specific tool.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The entire skill document is written as a mandatory policy in Chinese for all agents, with universal language such as '所有 Agent 必须遵守' and no indication that users may choose another language or locale. This effectively imposes a single language/locale behavior without opt-in, which matches the policy category for natural-language language/locale violations.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The entire skill is written as a universal instruction for '所有 Agent 工作都必须遵守' and contains no indication that language choice is optional or user-driven. Because it imposes a specific language/locale context across all agent work without opt-in or justification, it conflicts with the language/locale policy criteria.

Content

No source excerpt is available for this finding.