Back to skill

Security audit

Claim-to-Source Auditor

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed fact-checking workflow with no executable code, hidden install behavior, or evidence of exfiltration, though users should watch its CSV-saving and article-editing boundaries.

Installers should expect the skill to read article/source material, search for evidence, produce audit CSV/Markdown outputs, and possibly apply accepted factual corrections. Before use, decide where CSV and gold-set files should be saved, avoid feeding unnecessary confidential logs or drafts, and require explicit approval before any article file is modified.

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
  • 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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (5)

Natural-Language Policy Violations

Medium
Confidence
81% confidence
Finding
The manifest description defines activation using only Chinese trigger phrases and specifies Chinese status labels such as 通过/部分支持/缺证/错误/判断. The file does not clearly offer a user language choice for invocation or output, which can violate language/locale policy when imposed without opt-in.

Description-Behavior Mismatch

Medium
Confidence
97% confidence
Finding
The manifest description at L009-L010 limits the skill to auditing facts and sources and explicitly says it does not edit style or argument structure. However, the hard rule at L141 states that published articles may be revised directly when the user accepts a change, which expands the behavior from auditing into content modification.

Autonomous Decision Making

Medium
Category
Excessive Agency
Content
- A direct quote in an article that cannot be found in the reported source is a P0, even if the surrounding story is true.
- Near-miss numbers are still mismatches. An audit that reports 5810 when the source says 5817 must flag the difference.
- Cross-platform periods are high-risk. If one version reads "2024" and the other "2025", the different year is an error, not a paraphrase.
- Regulatory documents exist in specific jurisdictions. Never write "FDA, EU MDR, and China's regulations all say the same thing" without checking the governing text.

## Failure handling
Confidence
75% confidence
Finding
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Intent-Code Divergence

Low
Confidence
84% confidence
Finding
The architecture section describes the workflow as shadow-mode auditing and says not to modify the source article unless the user explicitly asks for revision. Later, L141 introduces direct revision behavior for published articles when the user accepts a change, creating an internal intent mismatch between read-only audit framing and edit-capable operation.

Missing User Warnings

Low
Confidence
90% confidence
Finding
The skill instructs saving audit results and verdicts to CSV as part of the default workflow, but it does not require explicit user consent before writing files. In environments with file-system or workspace access, this can cause unauthorized persistence of potentially sensitive article content, claims, internal data references, or prior audit history.

Static analysis

No suspicious patterns detected.