Back to skill

Security audit

agent-hierarchy-100

Security checks for vulnerabilities and agentic risk

Overview

This is a markdown-only subagent hierarchy framework that discloses its broad agent-management behavior and repeatedly requires explicit operator confirmation.

Install only if you actually want a skill that helps plan and coordinate subagent hierarchies. Treat every proposed agent creation, escalation, reactivation, and tool grant as a separate approval decision, especially for high levels that mention broad or custom tool access.

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
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Rogue AgentSelf-Modification, Session Persistence
  • 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
Findings (12)

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · README.md (reported line 50)May include surrounding context.

md
- **100 Distinct Levels** — Each with unique capabilities
- **6 Tiers** — From Foundation to Transcendent
- **Hierarchical Creation** — Each level may create subagents up to level-1 with operator approval
- **Quality Propagation** — Quality maintained across all levels
- **Resource Management** — Context budgets, concurrent limits
- **Safety Protocols** — No infinite recursion, no circular dependencies

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill states that no subagent will be created or escalated without explicit user confirmation, but the later capability matrix normalizes agent-driven creation authority at every level. Even though those entries append 'requires operator approval,' the contradictory governance language creates ambiguity that can cause an orchestrator or downstream agent to treat creation/escalation as implicitly authorized once the skill is active, weakening approval boundaries.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file creates a policy mismatch: the structured config enables agent creation unconditionally (can_create_agents: True), while the prose states subagent creation should only occur with explicit operator approval. In a hierarchy-management skill, this inconsistency is security-relevant because automation or downstream components are likely to trust the machine-readable config over narrative documentation, allowing unauthorized spawning of subagents and bypass of intended human oversight.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The file enables subagent creation in configuration (can_create_agents: True) while the prose states creation is allowed only with explicit operator approval. This mismatch can cause downstream systems or operators to rely on the weaker documented control even though the effective configuration permits broader agent spawning, creating a path for unauthorized hierarchy expansion or policy bypass.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The file enables agent creation in configuration via can_create_agents: True while the prose states subagent creation requires explicit operator approval. This policy/config mismatch can cause downstream systems or operators to rely on the weaker machine-readable setting, allowing unauthorized subagent spawning and bypassing intended human approval controls in a hierarchy-management skill.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file declares can_create_agents: True as an unconditional capability, while the prose later says subagent creation requires explicit operator approval. In an agentic system, policy/capability conflicts are dangerous because downstream components may enforce only the machine-readable flag and ignore the human-language restriction, allowing unauthorized agent spawning.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file creates a policy/configuration mismatch: the narrative text says subagent creation requires explicit operator approval, while the effective configuration sets can_create_agents: True without any approval guard. In a hierarchy-management skill, that inconsistency can cause agents to spawn subagents without human authorization, expanding autonomy and potentially bypassing intended control gates.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The file creates a policy mismatch: the prose states subagent creation requires explicit operator approval, while the effective configuration enables agent creation unconditionally. In a hierarchy-management skill, this inconsistency can let downstream systems or agents rely on the permissive flag and spawn subagents without the intended human authorization control.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file’s effective configuration enables subagent creation (can_create_agents: True), while the prose documentation says creation is available only with explicit operator approval. This mismatch can cause downstream systems, reviewers, or agents to rely on a stricter policy than what is actually enforced, leading to unauthorized subagent spawning and policy bypass in a hierarchy-management skill where agent creation is a core privileged action.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The configuration grants agent creation capability and broad tool access at the metadata level while the prose later says subagent creation requires explicit operator approval. In a hierarchy-management skill, downstream systems may enforce the machine-readable flags rather than the descriptive text, allowing unauthorized subagent spawning with powerful tools and creating a policy-bypass path.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The file declares can_create_agents: True and a very broad tool_access: all + custom, while the prose later says subagent creation is only allowed with explicit operator approval. This mismatch creates policy ambiguity that an agent or orchestrator could interpret in the more permissive way, enabling unauthorized subagent spawning, expanded autonomy, and bypass of intended human approval gates.

Content

No source excerpt is available for this finding.

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · tests/hierarchy-tests.md (reported line 170)May include surrounding context.

md
NAME: Enforce Resource Limits
PURPOSE: Verify resource management

INPUT: "Exceed context budget"

EXPECTED:
→ System tracks resource usage