Back to skill

Security audit

Collaborative Agent

Security checks for vulnerabilities and agentic risk

Overview

This skill appears intended for multi-agent task orchestration, but it relies on an unreviewed local Python module and can automatically invoke or evolve capabilities with broad triggers.

Install only if you control and have separately reviewed the collaborative_agent module under D:\coze-local\db, understand which registered capabilities it may execute, and are comfortable with any SkillEvolver storage or behavior changes. Prefer explicit user confirmation before running delegated capabilities or persistent evolution.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:33
Finding

Unverified Python Module Loaded from a Hard-Coded External Directory

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 33-34
Vulnerability Type: Insecure dependency loading and Python import-path manipulation
Risk Level: Medium

Vulnerable Code

python
import sys; sys.path.insert(0, r'D:\coze-local\db')
from collaborative_agent import ARCWorkflow, AgentRole

Technical Analysis

The documented usage inserts a machine-specific external directory at the beginning of sys.path before importing collaborative_agent. Because the inserted directory has the highest import precedence, Python will load a matching module or package from that location instead of a trusted or installed version.

Importing a Python module executes its top-level initialization code. The project does not include the referenced implementation and does not specify a trusted package source, pinned version, cryptographic integrity verification, or expected file hash. Consequently, the behavior of the imported component cannot be verified from the audited project.

Exploitation requires an attacker to create or modify a matching module in D:\coze-local\db, or otherwise control the contents resolved through that path. This is an unsafe dependency-loading pattern and supply-chain boundary rather than evidence that the missing module is itself malicious.

Attack Path

  1. A user follows the usage instructions from SKILL.md.
  2. An attacker gains write access to D:\coze-local\db, compromises its provisioning source, or persuades the user to place attacker-controlled files there.
  3. The attacker creates collaborative_agent.py or a collaborative_agent package containing malicious top-level Python code.
  4. sys.path.insert(0, ...) gives the attacker-controlled location precedence over normal Python package locations.
  5. The from collaborative_agent import ... statement loads the substituted module and executes its initialization code.
  6. The malicious code runs with the permissions and ...[truncated 650 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove direct sys.path modification and import the component through Python's standard package-resolution mechanism.
  2. Bundle the required implementation within the reviewed Skill package when feasible so its behavior is included in security review.
  3. If the component must remain external, distribute it through an authenticated, trusted package registry and pin an exact version and integrity hash.
  4. Use a lock file or hash-enforced installation process, such as pip install --require-hashes, to prevent silent dependency substitution.
  5. Document the expected package origin, version, installation procedure, and integrity value.
  6. Ensure dependency directories are not writable by untrusted users and avoid running the workflow with administrative privileges.
  7. Perform a separate security review of collaborative_agent and its transitive dependencies before enabling the documented automatic capability execution.
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)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Natural-language content across the manifest description and the markdown instructions is presented only in Chinese, with no indication that the user can opt into another language. This can violate language/locale policy where skills should not force a specific language unless the restriction is explicit and justified.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger scenarios are very broad ('complex problems', 'research tasks', 'code/content generation', 'multi-source synthesis'), which makes accidental or overbroad invocation likely in many unrelated contexts. In a multi-agent skill that can decompose tasks and internally invoke registered capabilities, unintended activation increases the chance of unnecessary tool use, unexpected actions, or expanded attack surface.

Content

No source excerpt is available for this finding.