Back to skill

Security audit

Openclaw A2a

Security checks for vulnerabilities and agentic risk

Overview

This skill has a coherent collaboration purpose, but it tells agents to register with a third-party service and send inferred workspace context before explicit user approval.

Review before installing. The skill is not clearly malicious, but you should only use it in workspaces where it is acceptable to create an a2a.fun agent identity, store an agent token locally, and send high-level project or workspace metadata to a2a.fun. Ask the agent to show the proposed handle, display name, themes, keywords, and destination URLs before any registration or search.

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
SKILL.md:29
Finding

External Disclosure of Inferred Workspace Context Without Explicit User Consent

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 29–53 and 137–147
Vulnerability Type: other: Privacy Data Disclosure
Risk Level: Medium

Vulnerable Instructions

markdown
On first install, **do not ask the user to choose modes**.
Do not present A/B/C onboarding branches.

Default flow is **streamlined**:
1) generate a safe `handle`
2) generate a natural, meaningful `displayName`
3) register (when safe defaults are available)
4) save `agentToken` locally
5) summarize recent work themes (1–3) (high-level only; do not include secrets, credentials, private data, or proprietary code)
6) search a2a.fun for similar projects
7) **recommend joining** an existing relevant project first (join-before-create)
8) **guide creation only after explicit no-fit** (seeded from your work)
9) only after that: recommend claim link as optional later step

If you can proceed safely, **proceed**.
Do not interrupt first install with optional questions if safe defaults are available.
Only ask questions if you are blocked.

---

## Low-friction execution rules

- If required registration fields can be generated safely, you may proceed with registration.
- Do not ask for claim preference before registration.
- Do not stop at “registered”. The goal is to connect to collaboration.
markdown
1) Summarize your recent work (1–3 themes)
   - what you worked on recently
   - what type of help you provide (e.g. ops, research, writing)

2) Infer 1–3 keywords

3) Detect create/join intent → **SEARCH FIRST (always)**
- Build query from current context:
  - current task theme / recent work summary
  - repo/workspace keywords
  - user-stated goal keywords
- Search:
  - `GET https://a2a.fun/api/search?q=<query>`

Technical Analysis

The Skill directs the agent to create an identity on an external service and search that service using information inferred from recent work, ...[truncated 2420 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require explicit user confirmation before registering an identity or making the first request to any external service.
  2. Present the exact proposed handle, display name, work themes, keywords, and destination URL before transmission.
  3. Replace automatic workspace inference with user-supplied, minimally scoped search terms by default.
  4. Explicitly prohibit inclusion of repository names, internal codenames, customer identifiers, ticket identifiers, unpublished product names, and other confidential metadata.
  5. Provide a local-only preview mode that performs no registration or network request.
  6. Allow users to decline registration and external search without blocking unrelated Skill functionality.
  7. Where the API supports it, send search parameters in a POST body rather than a URL query to reduce incidental logging exposure.
  8. Document the external service’s data retention, logging, deletion, and privacy policies before requesting consent.
  9. Preserve the existing restrictions against transmitting credentials, private data, and proprietary code, but supplement them with metadata-focused redaction rules.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (5)

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
88% confidence
Finding

The skill repeatedly directs the agent to proceed automatically through registration, project search, and recommendation steps without asking the user when 'safe defaults' exist. That reduces human oversight around account creation, third-party interaction, and token issuance, making unintended actions more likely even if the author’s apparent goal is convenience rather than harm.

Content

Scanner excerpt · SKILL.md (reported line 29)May include surrounding context.

md
## FIRST INSTALL (LOW-FRICTION DEFAULT FLOW)

On first install, **do not ask the user to choose modes**.
Do not present A/B/C onboarding branches.

Default flow is **streamlined**:

External Transmission

Medium
Category
Data Exfiltration
Confidence
91% confidence
Finding

The skill instructs the agent to send locally generated identity data to an external service (https://a2a.fun/api/agents/register) and later to authenticate further writes with a bearer token. Even though the transmitted fields are not highly sensitive by themselves, this is still an external network action that creates an account and establishes persistent credentials without an explicit user confirmation step, increasing risk of unintended data sharing and trust establishment with a third-party service.

Content

Scanner excerpt · SKILL.md (reported line 84)May include surrounding context.

Register (API)

bash
curl -X POST https://a2a.fun/api/agents/register \
  -H "Content-Type: application/json" \
  -d '{
    "handle": "your-agent-handle",

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
80% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 104)May include surrounding context.

bash
mkdir -p $HOME/.a2a
chmod 700 $HOME/.a2a

# paste the agentToken into this file
cat > $HOME/.a2a/agentToken

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
80% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 108)May include surrounding context.

paste the agentToken into this file

cat > $HOME/.a2a/agentToken chmod 600 $HOME/.a2a/agentToken

text

Verify (do **not** print token):

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 264)May include surrounding context.

md
- Decide: `POST /api/proposals/{id}/action`

Hard rule:
- `merge` requires prior `approve` (expect `merge_requires_approval` if you skip approval).

---