Back to skill

Security audit

summarize-pro

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because scanners found gaps between its advertised verification and what its checks enforce, which could give users misleading confidence in summaries.

Treat the advertised verification as limited: the scan reports that it does not establish factual support for every sentence or reliably qualify every action. Review summaries against their source, especially before relying on decisions or commitments. The verdict is suspicious (the Review category), not malicious; some references were unresolved, so scan coverage is incomplete.

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
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • 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
Findings (10)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill promotes narrow numeric checks as sentence-level provenance verification and directs delivery after checks pass, including for contracts and other high-risk documents. Unsupported claims and unlabeled commitments can therefore receive misleading assurance; however, delegating semantic summarization and intent routing to the agent is explicitly documented and is not itself a mismatch.

Content

No source excerpt is available for this finding.

citecheck can report PASS for wholly unsupported sentences, contrary to the claimed sentence-level provenance guarantee.

High
Category
Excessive Agency
Confidence
100% confidence
Finding

citecheck checks only selected numeric and date tokens against the entire source, so fabricated sentences without those tokens pass, even with zero tokens checked. Reusing a source number in a false claim also passes, allowing misleading verification of summaries that users may rely on for consequential decisions.

Content

No source excerpt is available for this finding.

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

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.

L23 instructs contextual correction of transcription errors, conflicting with the manifest's requirement to flag ambiguity rather than resolve it.

Medium
Category
Excessive Agency
Confidence
92% confidence
Finding

The manifest promises that nothing absent from the source will be introduced and that ambiguity will be flagged rather than resolved. L23 instead directs the summarizer to infer a correction from context; disclosing that inference preserves transparency but still permits an unverified interpretation to enter the summary.

Content

No source excerpt is available for this finding.

The complex minutes example presents conditional and unconfirmed items as ordinary action items, contrary to the manifest's commitment-grading discipline.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The example models a conditional release date as an ordinary action-table entry without preserving its prerequisite there, and assigns work without distinguishing assignment from acceptance. Because agents may imitate examples, this undermines the skill's integrity safeguards and can make conditional or unconfirmed work appear ready for execution.

Content

No source excerpt is available for this finding.

Third-party task assignments are classified as confirmed commitments without evidence that the assignee accepted.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

COMMIT_COMMITTED matches third-party assignments such as '由某某负责' without checking whether the named person accepted, and this match takes precedence over the unconfirmed category. This can incorrectly mark ownership as committed and contribute to a clean verdict, undermining the tool's central distinction between assignment and acceptance.

Content

No source excerpt is available for this finding.

selfcheck does not enforce the per-item strength labels its comments declare mandatory.

Medium
Category
Excessive Agency
Confidence
100% confidence
Finding

selfcheck accepts one matching label anywhere in a block instead of validating every item, and any occurrence of '原文未提及' bypasses that block's label check. An incomplete or manipulated summary can therefore pass the advertised strength-label gate while leaving consequential decisions or tasks unqualified.

Content

No source excerpt is available for this finding.

Dynamic import via __import__()

Medium
Category
Dangerous Code Execution
Confidence
75% confidence
Finding

Dynamic import() can load arbitrary modules at runtime, bypassing static analysis and potentially importing malicious code.

Content

Scanner excerpt · tests/run_golden.py (reported line 98)May include surrounding context.

python
body = sp.build_skeleton(fmt, facts)

    if exp.get("skeleton_has_numbered_lines") and not __import__("re").search(r"^\d+\.\s", body, __import__("re").M):
        fails.append("骨架缺少编号行")
    if "bullets_count_between" in exp:
        import re

Dynamic import via __import__()

Medium
Category
Dangerous Code Execution
Confidence
75% confidence
Finding

Dynamic import() can load arbitrary modules at runtime, bypassing static analysis and potentially importing malicious code.

Content

Scanner excerpt · tests/run_golden.py (reported line 98)May include surrounding context.

python
body = sp.build_skeleton(fmt, facts)

    if exp.get("skeleton_has_numbered_lines") and not __import__("re").search(r"^\d+\.\s", body, __import__("re").M):
        fails.append("骨架缺少编号行")
    if "bullets_count_between" in exp:
        import re

Broad keyword triggers can activate the summarizer for requests that do not involve summarization.

Low
Category
Trigger Abuse
Confidence
90% confidence
Finding

The trigger list includes standalone terms such as "timeline", "action items", and "comparison table" without requiring an intent to summarize existing source material. These terms also occur in planning, task-management, and original-content requests, and the manifest provides no exclusion conditions to distinguish those uses.

Content

No source excerpt is available for this finding.