Back to skill

Security audit

war-room

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this skill suspicious because scanners report GitHub publishing by default and external model sharing that may exceed its stated expert-review purpose; confidence is medium.

Consider the reported risk that decision details may be shared with external model providers and published to GitHub Discussions by default. The Skillspector scan was partial: it scanned all nine listed components, but three local path references were unresolved, so evidence is incomplete. The AIG scan completed with no findings, which does not negate the separate Skillspector findings. The verdict is suspicious, not malicious.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (15)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied command performs a GitHub mutation rather than deliberation or expert review. Although the comment identifier suggests it may refer to a panel synthesis, marking it as an accepted answer changes shared external state. This capability is not disclosed by the declared purpose and is more than an internal implementation detail.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The chunk performs an external write operation rather than expert-panel deliberation. Nothing in the description establishes GitHub discussion management as a supporting function, and no permissions are declared. The required GraphQL variables are not supplied in this snippet, so successful execution is not established, but the intended mutation is clear.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The documented workflow directs the agent to create a Discussion and post summaries using existing GitHub credentials, with an opt-out rather than affirmative authorization. The integration is disclosed in the body, but that disclosure does not safely authorize external writes containing potentially confidential deliberations.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill makes authenticated publication the default unless the user explicitly declines, although requesting expert deliberation does not inherently authorize remote publication. This can disclose sensitive strategic decisions and summaries to everyone with access to the selected repository, potentially including the public.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The supplied chunk performs a GitHub GraphQL metadata lookup rather than expert-panel deliberation. Repository Discussions access is not explained by the declared purpose, so it constitutes an undeclared resource integration. The code shown only reads metadata; it does not publish discussions or modify the repository.

Content

No source excerpt is available for this finding.

The skill defaults to publishing decision content to GitHub, beyond its stated expert-deliberation purpose.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The module explicitly defaults to publishing and uses authenticated GraphQL mutations to transmit decision context, rationale, and subsequent phase summaries. It neither requires affirmative authorization nor checks the destination audience or sensitive content, creating a concrete disclosure risk despite its stated publishing purpose.

Content

No source excerpt is available for this finding.

External-model delegation lacks a privacy warning about sharing user context and source files with third-party providers.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The skill passes all brainstorming context into a workflow that delegates to multiple external model providers, without specifying a pre-transfer privacy review or data minimization. Although multi-model delegation is an advertised feature, that does not establish permission to send confidential source files or unrestricted user context to every provider.

Content

No source excerpt is available for this finding.

The skill goes beyond convening an expert panel by requiring publication of decisions and deliberation summaries to GitHub Discussions by default.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The mandatory-by-default publishing workflow crosses from private decision support into authenticated external disclosure. Because the skill targets high-stakes architecture and strategy, its decisions and phase summaries may contain particularly sensitive information, and an opt-out prompt does not ensure informed authorization.

Content

No source excerpt is available for this finding.

Decision details and phase summaries are published externally by default unless the user explicitly opts out.

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

The instructions proceed with external publication unless the user says no, rather than requiring an affirmative grant. Decision rationale and intelligence or red-team summaries may expose proprietary plans or security weaknesses beyond the audience intended for the original deliberation.

Content

No source excerpt is available for this finding.

The deliberation workflow requires publishing the finalized decision to GitHub Discussions by default, exceeding the manifest's expert-review purpose.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The team workflow requires publishing the finalized decision by default, extending deliberation into an external write without affirmative permission. Sensitive high-stakes decisions could consequently reach an unintended repository audience.

Content

No source excerpt is available for this finding.

Makes external publication of the synthesized decision mandatory by default, without requiring affirmative user authorization.

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

The instruction that the user can decline with 'n' treats publication as authorized unless refused. That is an unsafe consent boundary for sending potentially confidential synthesized decisions outside the local session.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% 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 · modules/deferred-capture.md (reported line 60)May include surrounding context.

md
## Behavior Rules

- Capture is automatic: do not prompt the user for
  confirmation.
- Capture every qualifying COA in a single pass before
  closing the session.

The workflow sends user context to external model experts without a privacy warning.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

Phase 1 assigns user problem statements and supplied context files to external Qwen and Gemini experts without specifying a privacy or authorization gate. External expertise is expected in this workflow, but unrestricted transfer of sensitive files can still violate confidentiality or provider-use restrictions.

Content

No source excerpt is available for this finding.

Sealed serialization exposes a deterministic attribution hash that permits recovery of expert identities.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Sealed serialization exposes SHA-256 of the unsalted role:model string, whose possible values come from a small, largely known panel configuration. A reviewer can enumerate those combinations and recover attribution, defeating the anonymity intended to prevent model-preference bias during high-stakes review.

Content

No source excerpt is available for this finding.

The advertised tamper check does not verify deliberation history or graph integrity.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The advertised session verifier only checks each node's content and attribution against its truncated identifier; it does not authenticate parent links, phases, rounds, timestamps, node completeness, or the session root. History can therefore be altered or nodes removed while verification still succeeds, undermining the deliberation audit trail.

Content

No source excerpt is available for this finding.