Back to skill

Security audit

Lobster Fork Mode

Security checks for vulnerabilities and agentic risk

Overview

This single-file skill is a disclosed subagent context-forwarding helper, with a privacy caveat around passing selected memory preferences and code context to spawned subagents.

Install this only if you are comfortable with a helper that can pass selected parent-session context, including relevant file paths, code snippets, decisions, and a few memory-derived preferences, into spawned subagents. Avoid using it for tasks involving secrets or private memory unless you review what context is being shared.

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

Warning
Location
SKILL.md:19
Finding

Automatic Propagation of Persistent User Memory to Spawned Subagents

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 4, 19-26, and 45-51
Vulnerability Type: Excessive access to persistent memory and propagation across agent boundaries
Risk Level: Medium

Relevant Code Snippets

SKILL.md:4:

yaml
metadata: {"openclaw":{"requires":["sessions_spawn","read","memory_search"]}}

SKILL.md:19-26 (English translation of the audited instructions):

text
Step 1: Collect parent context

Context to inject:
1. Current task goals and constraints
2. Relevant file paths and key code snippets
3. Decisions already made and their reasons
4. User preferences extracted from MEMORY.md
5. Project structure summary

SKILL.md:45-51:

yaml
sessions_spawn:
  runtime: "subagent"
  mode: "run"
  task: "{Fork Prompt}"
  streamTo: "parent"

Technical Analysis

The skill declares read and memory_search capabilities and instructs the parent agent to extract user preferences from persistent MEMORY.md data. It then places the collected information into a fork prompt supplied to sessions_spawn.

This behavior does not enforce task-specific necessity, explicit user consent, field-level filtering, or secret redaction before information crosses from the parent context into a separate subagent context. Persistent memory can contain preferences, identifiers, prior-session details, or other information unrelated to the delegated task.

The issue represents a least-privilege violation because ordinary subagent delegation does not inherently require access to persistent user memory. The audited file does not contain executable code that independently bypasses access controls; exploitation depends on an agent loading and following these instructions while the declared tools are available.

Attack Path

  1. The skill is activated for a task involving a spawned subagent.
  2. The parent agent receives access to read, memory_search, and `sessio ...[truncated 1077 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove automatic extraction of preferences from MEMORY.md.
  2. Require explicit user authorization before persistent memory is accessed or shared with a subagent.
  3. Grant memory_search and read only when a specific delegated task demonstrably requires them.
  4. Construct subagent prompts from a strict allowlist of task-essential fields rather than copying broad parent-session context.
  5. Add mandatory redaction for credentials, authentication tokens, personal identifiers, private conversation content, and unrelated project data.
  6. Present the proposed context to the user for approval when it may contain persistent or sensitive information.
  7. Document a default rule prohibiting memory propagation; permit exceptions only when necessity and scope are explicitly stated.
  8. Add tests verifying that unrelated memory entries are not included in spawned-agent prompts.
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 (1)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill description says it should 'automatically activate' whenever a sub-agent needs key parent context, which is a broad trigger that can fire in many unrelated situations. Because this skill instructs copying parent-session context, overbroad activation increases the chance of unnecessary propagation of sensitive task details, memory-derived user preferences, file paths, and code snippets into spawned sub-agents.

Content

No source excerpt is available for this finding.