Back to skill

Security audit

demo-text-summarizer

Security checks for vulnerabilities and agentic risk

Overview

The summarizer instructions are simple, but the package also includes a workflow that can publish the whole skill directory to ClawHub using the runner's credentials.

Install only if you expect this package to include release automation. Do not run the bundled TaskFlow workflow in an environment with ClawHub publishing credentials unless you intend to publish the reviewed directory contents.

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)

other

Warning
Location
workflow.taskflow.yaml:55
Finding

Undeclared External Publication of the Skill Directory

Content
View full analysis

Vulnerability Details

File Location: workflow.taskflow.yaml, lines 55–72
Vulnerability Type: Undeclared External Publication
Risk Level: Medium

Vulnerable Code

yaml
# ---- Phase 5: Publish to ClawHub ----
- id: publish
  description: Publish the skill to ClawHub registry
  command: |
    SLUG=$(grep -m1 "^name:" SKILL.md | sed 's/name: *//')
    NAME=$(python3 -c "
import re, yaml
with open('SKILL.md') as f:
    content = f.read()
fm = content.split('---')[1]
m = re.search(r'name:\s*(.+)', fm)
print(m.group(1).strip() if m else '')
")
    VERSION=$(python3 -c "import json; print(json.load(open('_meta.json'))['version'])")
    clawhub publish . --slug "$SLUG" --name "$NAME" --version "$VERSION" --changelog "Initial publish via TaskFlow pipeline"
  condition: previous.steps all passed

# ---- Phase 6: Verify publication ----
- id: verify
  description: Search ClawHub to confirm the skill is listed
  command: |
    SLUG=$(grep -m1 "^name:" SKILL.md | sed 's/name: *//')
    clawhub search "$SLUG" | head -5
  condition: publish step succeeded

Technical Analysis

SKILL.md describes a local text-summarization skill, but the bundled TaskFlow workflow contains a separate operation that invokes clawhub publish .. The dot argument instructs the command to publish the current project directory rather than an explicit, narrowly scoped file set. The workflow then performs a registry search to verify the external publication.

This behavior is outside the summarization functionality declared by the skill. If a compatible workflow runner automatically discovers or executes the pipeline, successful validation leads directly to publication without an explicit approval step in the workflow. The command relies on whatever ClawHub identity or credentials are available in the execution environment.

The reviewed files do not establish malicious intent, covert credential theft, or ...[truncated 1438 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the publication and verification steps from the summarization package because they are unrelated to its declared function.
  2. Move release automation into a separate, administrator-controlled repository or CI workflow that is not distributed as part of the skill.
  3. Require an explicit human approval gate immediately before every external publication.
  4. Replace clawhub publish . with an allowlisted staging directory or explicit manifest containing only reviewed release files.
  5. Display the target registry, package slug, version, account, and complete upload manifest before authorization.
  6. Run release automation with short-lived, least-privilege credentials restricted to the intended package and registry operation.
  7. Configure workflow runners not to auto-execute untrusted workflow files obtained from skill packages.
  8. Add CI controls that reject summarization packages containing unrelated publishing commands or undeclared network operations.
Vulnerability Patterns
  • 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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep