Back to skill

Security audit

Consensus Commons

Security checks for vulnerabilities and agentic risk

Overview

This skill is not clearly malicious, but it can publish and lock governance-style decisions externally while overstating its validation and review guarantees.

Review before installing. Use only mock mode or non-sensitive test topics unless you are comfortable sending the content and station credentials to Spacebase1. Do not rely on LOCKED, CHP, audit trail, confidence, or specialist-agent labels as proof that real independent validation occurred.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
council.py:340
Finding

CHP Validation Is Bypassed and Supported Decisions Are Unconditionally Locked

Content
View full analysis
Remediation
View remediation
CHALLENGED -> UNDER_REVIEW -> VALIDATED -> LOCKED`. 6. Remove direct `PROVISIONAL -> LOCKED` and `CHALLENGED -> LOCKED` transitions. 7. Check and handle every `lock_intent()` return value rather than continuing after a failed transition. 8. Require recorded human approval for high-impact or high-value decisions. 9. Add tests proving that missing or failed gates prevent validation and locking. ]]>

T09 · Insecure Skill Coding Practices

Warning
Location
council.py:39
Finding

Hard-Coded Agent Conclusions Are Presented as Independent Multi-Agent Analysis

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
routing.py:76
Finding

Bypassable Sensitive-Data Filter Can Publish Confidential Content to an External Service

Content
View full analysis
0.2: return RouteDecision( role="reject", is_supported=False, reason="Intent references private or confidential data. " "This topic requires a private council and cannot be " "processed in the public commons.", confidence=reject_score, tags=["private", "rejected"], ) ``` The content is subsequently sent to the external Spacebase1 service: ```python # client.py:288-299 async def _send_itp(self, verb: str, body: dict[str, Any]) -> dict[str, Any]: """Send an ITP act and return the parsed response.""" http = self._ensure_connected() envelope = { "verb": verb, "sender": self.agent_name, "timestamp": time.time(), "body": body, } resp = await http.post(ITP_ENDPOINT, json=envelope) resp.raise_for_status() return resp.json() ``` ```python # client.py:329-361 async def post( self, content: str, parent_id: str | None = None, payload: dict[str, Any] | None = None, ) -> Intent: result = await self._send_ ...[truncated 2962 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (11)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied code is only a narrow validation helper for CHP gates. While it loosely aligns with the declared consensus-hardening theme, it does not implement the broader advertised skill behavior such as intent routing, multi-agent consensus building, contrarian review, governance workflow orchestration, or a lock state machine with multiple states. The description substantially overstates the implemented functionality, so this is a mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The declared description presents a full decision-orchestration system with consensus hardening, specialist routing, adversarial review, governance workflow, and integration/runtime modes. The supplied code is only a stub module containing enums and dataclasses (TurnResult, Workflow) plus minimal aggregate calculations. Its lock logic has only a few states (FAILED, LOCKED, PROVISIONAL) rather than a 5-state machine, and there is no actual orchestration, routing, review, panel selection, audit pipeline, or external integration. This is a material description-to-behavior mismatch, not merely an incomplete supporting component, because the code chunk’s primary function is passive schema/aggregation support rather than the described council system behavior.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The declared description presents a complete multi-agent council system with consensus hardening, adversarial review, governance workflow validation, and audit-trail production. The supplied code chunk only performs intent routing: it normalizes text, scores keyword matches, rejects private topics, and returns a RouteDecision with role and agent list. While this routing supports the declared system at a component level, the description as given overstates what this code chunk actually does. The primary implemented purpose here is intent classification and agent-panel assignment, not the broader deliberation and consensus machinery.

Content

No source excerpt is available for this finding.

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
80% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · chp.py (reported line 53)May include surrounding context.

python
@property
    def is_locked(self) -> bool:
        """Valid and no warnings (strict mode)."""
        return self.is_valid and len(self.warnings) == 0

    def to_dict(self) -> dict[str, Any]:

Lp1

High
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The file implements real outbound HTTP connectivity via httpx to fixed external endpoints, but the static finding indicates this network capability is not covered by declared permissions. In an agent skill, undeclared network access is dangerous because it expands the skill's effective authority, can exfiltrate data such as prompts or tokens, and may surprise operators who expected offline-only or permission-bounded behavior. The multi-agent governance context increases risk because sensitive deliberation data and station credentials may flow through this client.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The invocation guidance is extremely broad ('use whenever the user needs ... any structured decision-making process'), which can cause the agent framework to route many unrelated or sensitive requests into this skill. In a governance/finance context, overbroad auto-invocation increases the chance that confidential, regulated, or high-stakes decisions are processed by a skill whose actual guarantees may be weaker than advertised.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The general trigger keywords ('should', 'decide', 'recommend', 'evaluate', etc.) are so common that ordinary conversation is likely to activate this skill unintentionally. That creates a prompt-routing vulnerability where benign user queries may be diverted into a tool-enabled deliberation workflow, increasing exposure of user data and the chance of inappropriate automation in sensitive contexts.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The mock client docstring says operations are deterministic and can be replayed identically. However, post and post_child generate UUID-based IDs and use the current wall-clock time, so repeated runs will not produce identical artifacts or logs.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The docstring states that all writes are signed, but the implementation only uses bearer-token authenticated HTTP requests and never applies a request signature or DPoP-style proof. This mismatch is security-relevant because callers may rely on stronger sender authenticity, replay resistance, or non-repudiation guarantees that are not actually present, leading to unsafe deployment assumptions. In a consensus/audit-trail system, overstating integrity protections is especially dangerous because governance outputs may be treated as trustworthy evidence.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This code performs a network request that transmits credential material (station_token) in both the Authorization header and JSON body. For code files, such transmission should have some visible disclosure such as a confirmation prompt, user-facing log/print, or explicit warning in surrounding documentation/comments; here the only note is an internal docstring/comment and the user-facing log occurs only after success, not as a warning before transmission.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The runner automatically transitions the root intent to CHALLENGED and then LOCKED based solely on internal agent outputs, with no explicit human confirmation, policy gate, or caller-provided approval step. In a governance or decision-making skill, this can prematurely finalize decisions, suppress further review, and create unsafe or unauthorized outcomes if the routed intent or agent analysis is wrong, manipulated, or incomplete.

Content

No source excerpt is available for this finding.