Back to skill

Security audit

sz-ink

Security checks for vulnerabilities and agentic risk

Overview

This rewriting skill needs review because required searches can expose document details, and its safeguards can miss changes to facts or meaning.

Keep confidential text out of external searches unless you explicitly approve sanitized queries. Manually verify facts, numbers, and negations even when checks pass. Review or remove saved drafts and profiles under .agent/harness/ when finished.

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
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • 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 (8)

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Few-shot rewrite examples introduce unsupported facts despite the manifest's semantic-preservation requirement.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The official-document examples introduce unsupported operational facts, deadlines, and quantitative results, including a 12% cost reduction, while presenting the transformations as fact-preserving. Because these examples are explicitly injected into the rewriting context, they can steer the agent into fabricating authoritative-looking statements; downstream review reduces but does not eliminate this output-integrity risk.

Content

No source excerpt is available for this finding.

Mandatory searches of input details lack a privacy warning.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow mandates external searches of detailed input material without a sensitivity check or user authorization gate. Since accepted inputs include work reports, letters, and other potentially private documents, search queries could disclose confidential details to a search provider even when the purpose is merely checking originality.

Content

No source excerpt is available for this finding.

Rewrite prescriptions introduce unsupported facts despite the manifest's semantic-fidelity requirement.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

The rhythm-rewriting example changes completion in the third quarter into work lasting one quarter and invents three rounds of research. These prescriptions guide actual rewriting, so the example encourages unsupported factual additions despite the semantic-fidelity requirement; preserving original anchors alone would not reliably catch these additions.

Content

No source excerpt is available for this finding.

The prescription uses blacklist-triggered rewriting despite the manifest's claim that rewriting does not rely on forbidden-word lists or mechanical synonym replacement.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The manifest explicitly states “不靠禁词表和同义词硬换,” but this card makes a single vocabulary-list match sufficient for diagnosis and supplies direct substitutions such as “赋能→帮/让.” Although quotation and register exemptions provide some safeguards, the prescribed trigger still relies on listed words rather than requiring contextual evidence that their use is inappropriate.

Content

No source excerpt is available for this finding.

The official-document workflow enforces a banned-word gate despite the manifest explicitly rejecting reliance on banned-word lists.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

The manifest describes context-sensitive rewriting that does not rely on banned-word lists ("不靠禁词表和同义词硬换"). L31 instead requires unconditional blocking whenever listed jargon appears in a formal document ("出现即阻断"), making word-list membership a decisive criterion rather than a contextual diagnostic.

Content

No source excerpt is available for this finding.

L08 promises that negation counts must not decrease, but L53 permits a decrease of one without reporting drift.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The tolerance silently accepts the loss of one negation, so a sentence such as “不建议使用” becoming “建议使用” can receive PASS when no other check detects it. Independent semantic review remains required and mitigates the risk, but the mechanical safeguard still fails to flag a potentially complete reversal of the author's position.

Content

No source excerpt is available for this finding.

L004 claims thresholds are read from rules/metrics.yaml, but the script never loads that file or applies its thresholds.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

The module documentation explicitly identifies rules/metrics.yaml as the threshold source. However, run() at L200–L224 only calculates raw metrics using local lexicons, and the CLI at L276–L291 prints those results without loading any threshold configuration; the documented threshold integration is absent.

Content

No source excerpt is available for this finding.