Back to skill

Security audit

Lobster Coordinator

Security checks for vulnerabilities and agentic risk

Overview

The skill is a disclosed multi-agent coordinator, but it explicitly tells child agents not to inherit parent denial rules, which creates a real permission-boundary concern.

Review this skill carefully before installing. Its coordination features are useful for complex tasks, but only install it in environments where child-agent permissions are enforced by the platform and cannot bypass parent denial rules; avoid using it with broad read, write, exec, or network tools unless you explicitly approve the scope.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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 (1)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:55
Finding

Child Agents Explicitly Bypass Parent Denial Rules

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 55–59
Vulnerability Type: Safety constraint bypass during sub-agent delegation
Risk Level: High

Vulnerable Code

markdown
### Permission isolation rules
- Inherit the parent Agent's alwaysAllow rules
- **Do not** inherit alwaysDeny rules (sub-Agents may require different permissions)
- Independent tool allowlist (specified in the Agent definition)
- High-risk operations require separate user confirmation

The original source text is written in Chinese; the excerpt above is a faithful English translation of the complete affected section.

Technical Analysis

The skill explicitly instructs delegated child agents to inherit the parent agent's allowed permissions while not inheriting its alwaysDeny restrictions. This creates an asymmetric permission model that preserves grants but removes prohibitions.

This behavior violates least privilege and constitutes an instruction-level safety constraint bypass. The issue is especially significant because the agent-definition example at lines 35–42 permits powerful tools such as filesystem read/write access, command execution, and web search:

markdown
---
description: Agent description
tools: [read, write, exec, web_search]  # Tool allowlist
model: zai/glm-5.1
effort: thorough
background: true
---
# Agent instructions...

A statement that high-risk operations require user confirmation does not restore the discarded denial rules. The document does not define an enforceable mechanism that guarantees confirmation before each prohibited operation.

Attack Path

  1. A complex task activates the skill's sub-agent delegation workflow.
  2. The coordinator constructs or loads an agent definition containing tools such as read, write, exec, or web_search.
  3. The coordinator creates the child through sessions_spawn.
  4. Under the documented permission rules, the child inherits the parent's alwaysAllow rules but not its `alwaysDe ...[truncated 1090 chars]
Remediation
View remediation

Remediation Suggestions

  1. Require every child agent to inherit all parent alwaysDeny rules without exception.
  2. Calculate child permissions as an intersection of the parent's effective permissions and a narrowly scoped child allowlist.
  3. Never allow skill instructions or agent definitions to weaken platform-, administrator-, or user-defined denials.
  4. If exceptional access is legitimately required, use an enforceable elevation mechanism that:
    • Identifies the exact tool and operation.
    • Limits access to specific resources and arguments.
    • Explains the reason and security impact to the user.
    • Requires explicit user approval before elevation.
    • Applies only to one task or operation.
    • Records the approval and subsequent privileged actions.
  5. Default child agents to minimal capabilities and avoid broad combinations such as unrestricted read, write, exec, and network access.
  6. Enforce high-risk confirmations in the runtime authorization layer rather than relying on natural-language instructions.
  7. Add tests verifying that operations denied to the parent remain denied in every spawned child, background worker, and persistent session.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (2)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill description states it will auto-activate when tasks become complex, but the trigger is broad and subjective. In a coordinator skill with permissions to spawn subagents and send inter-session messages, ambiguous activation can cause unnecessary delegation, expanded tool use, and unintended persistence/communication without clear user intent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The description is written to operate in Chinese without any indication that language selection follows the user's preference. This can degrade transparency and informed consent by causing the agent to switch languages unexpectedly, increasing the risk that users misunderstand delegation status, permissions, or outputs from subordinate agents.

Content

No source excerpt is available for this finding.