ClawHub
Back to skill

Security audit

literature-review-paper-screener

Security checks for vulnerabilities and agentic risk

Overview

This skill clearly discloses its paid cloud workflow and requires approval, though a validation bug can block legitimate results.

Expect a separately installed CLI, an account, and paid third-party screening. Review the destination, transferred evidence, and quoted cost before approving each run. Use a dedicated output directory and a reviewed CLI version. The excluded-paper validation bug should be fixed before relying on complete review deliverables; do not omit exclusions to obtain a pass.

A.I.G

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

SkillSpector

Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (5)

The audit command supplies a manifest outside the declared file-access scope.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The audit command passes ../store-manifest.json to the validator despite the declared read scope covering bundled files and run artifacts, and the read inventory not mentioning a manifest. Depending on the working directory, this directs the helper to read a parent-directory file outside the intended boundary; the supplied content does not establish exfiltration or how the manifest is processed.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% 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.

Content

Scanner excerpt · SKILL.md (reported line 441)May include surrounding context.

md
## Operating rules

- **Human-in-the-loop (global, mandatory — covers every high-impact action).** This Skill never
  auto-executes actions that spend money, send data to a third party, or irreversibly change
  files. Every cloud submission, every data handoff, and every file overwrite requires an
  explicit confirmation from the user in the current conversation, after the relevant numbers
  (platform, fee, paper count) have been shown. Silence, a previous approval, or an earlier

Autonomous Decision Making

Medium
Category
Excessive Agency
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.

Content

Scanner excerpt · SKILL.md (reported line 447)May include surrounding context.

md
(platform, fee, paper count) have been shown. Silence, a previous approval, or an earlier
  choice is NOT a confirmation.
- **Ask the user which platform before every cloud run — including test and verification runs.**
  Never default to the currently active profile without asking (step 1.5), and never skip the
  platform + fee + count confirmation ceremony (step 8) for any run, test or not. The confirmation
  must come from the user after the numbers have been shown: silence, a previous run's approval, or
  an earlier platform choice is not a confirmation, and if the user does not respond you do not

Autonomous Decision Making

Medium
Category
Excessive Agency
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.

Content

Scanner excerpt · SKILL.md (reported line 451)May include surrounding context.

md
platform + fee + count confirmation ceremony (step 8) for any run, test or not. The confirmation
  must come from the user after the numbers have been shown: silence, a previous run's approval, or
  an earlier platform choice is not a confirmation, and if the user does not respond you do not
  submit. A run that spends money without asking is a rule violation, not an internal shortcut.
- **Handoff content and delivered outputs follow the language rule in step 5** (English by
  default, or the assignment-required language when one was specified). Literature content always
  keeps its original language — cite a paper in the language it was published in, and never

The validator requires a Paper Sheet entry for excluded-paper rows despite documenting that excluded papers must never appear there.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Every row must contain exactly one paper_sheet entry, including rows whose sole screening decision is exclusion, but audit() rejects any excluded record appearing in paper_sheet and any entry outside the screened pool. Consequently, a correctly represented excluded-paper row cannot pass validation, creating a validation denial of service for legitimate screening results and pressure to omit exclusions or misclassify papers; this undermines the literature-review skill's completeness and auditability.

Content

No source excerpt is available for this finding.