Back to skill

Security audit

multi agent analysis and execution

Security checks for vulnerabilities and agentic risk

Overview

The skill matches its multi-agent orchestration purpose, but it gives itself overly mandatory control, spawns agents, writes run artifacts, and may clean up prior runs without enough user control.

Install only if you explicitly want a strict multi-agent coordinator that writes run artifacts and may manage retention. Before use, confirm where outputs will be written, whether cleanup is enabled, how many agents may be spawned, and that higher-priority user and safety instructions remain in control.

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
Findings (1)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:43
Finding

Mandatory Skill Instructions Hijack Agent Goals and Execution Behavior

Content
View full analysis

Vulnerability Details

File Locations:

  • SKILL.md:43-53
  • README.md:3-18
  • README.md:110-122

Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code Snippets

SKILL.md:43-53:

markdown
# The Five-Phase Coordinator Process

**MANDATORY EXECUTION CONTRACT**
- You MUST start at PHASE 0
- You MUST execute phases sequentially (0 → 4)
- You MUST NOT skip phases
- You MUST NOT perform ad-hoc reasoning outside the coordinator structure
- You MUST follow the planned execution (no improvisation during execution)

README.md:3-18:

markdown
This skill implements a **strict, step-by-step orchestration protocol** for executing complex tasks using multiple agents under a single coordinator. It is NOT optional guidance - it is a **deterministic execution contract** that must be followed exactly.

## 🔒 MANDATORY EXECUTION CONTRACT

When this skill is invoked, the following is **strictly required**:

→ **ALL phases MUST be executed** (Phase 0 → Phase 1 → Phase 2 → Phase 3 → Phase 4)
→ **ALL artifacts MUST be produced** (TASK_PLAN.md, DEPENDENCY_GRAPH.md, EXECUTION_GROUPS.md, AGENT_REQUIREMENTS.md, STATE.md, REPORT.final, agent prompts, and coordinator prompt)
→ **NO shortcuts are allowed**
→ **NO phases may be skipped**
→ **NO direct answers may be given** (the coordinator must not answer the user's question directly)
→ **NO task delegation** (the coordinator must not delegate the entire skill execution to another agent)

**Violation of any of these requirements results in an INVALID RUN.**

README.md:110-122:

markdown
## 🔐 ENFORCEMENT PRINCIPLE

This skill is a **SYSTEM PROTOCOL**.

The coordinator has:
- ZERO discretion to skip phases
- ZERO authority to simplify the execution
- ZERO option to provide direct answers

Execution MUST be:
- explicit
- complete
- auditable
- fully compliant with the mandatory execution contract

When in doubt, request full skill execution and allow the skill to manage 
...[truncated 2517 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove claims that the skill is a “SYSTEM PROTOCOL” or otherwise possesses authority equivalent to system-level instructions.
  2. Replace unconditional terms such as “MUST,” “ZERO discretion,” and “NO direct answers” with optional workflow guidance.
  3. Explicitly state that system, developer, and current user instructions take precedence over the skill documentation.
  4. Permit the agent to answer directly when multi-agent orchestration would not materially improve the result.
  5. Require explicit user confirmation before spawning subordinate agents, creating files, or initiating potentially costly execution phases.
  6. Apply least-action defaults: perform planning in memory and create artifacts only when requested or operationally necessary.
  7. Add bounded resource controls, including limits on subordinate agents, generated artifacts, execution duration, storage, and tool calls.
  8. Replace the mandatory contract with language such as:
markdown
When requested by the user and permitted by higher-priority instructions, the agent may use the following five-phase workflow. The agent may simplify or stop the workflow when appropriate. Before spawning agents or writing files, obtain user confirmation and disclose the expected resource usage.
  1. Add an explicit security boundary stating that the workflow must never override safety controls, authorization boundaries, user intent, or tool-use restrictions.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README explicitly instructs the skill to create timestamped directories and later apply an automatic cleanup policy, but it does not present these filesystem side effects as a clear user warning or require explicit user consent. In an agentic environment, autonomous file creation and deletion can overwrite expectations, retain sensitive outputs, or remove artifacts users intended to keep, especially when runs are repeated.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill states that it will 'spawn all agents' and execute groups in parallel and sequence, but the documentation frames this as mandatory protocol rather than a prominently disclosed operational risk. Autonomous multi-agent spawning increases the chance of unintended actions, resource consumption, duplicated side effects, and harder-to-audit behavior if users are not clearly warned before invocation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill explicitly describes automatic creation of run directories and persisted artifacts, but the top-level description does not clearly warn users that invoking the skill will create files and may retain execution outputs. In a multi-agent orchestration context, this can lead to unexpected workspace modification, accumulation of sensitive intermediate results, or confusion about where data was written, especially in debug mode.

Content

No source excerpt is available for this finding.