Back to skill

Security audit

Claw Crony

Security checks for vulnerabilities and agentic risk

Overview

The skill has a coherent A2A setup purpose, but its install, credential, and network instructions are too broad and risky for automatic installation.

Install only after reviewing and pinning the plugin source and dependencies, replacing curl-to-shell installation with a safer package-manager flow, binding the gateway to a trusted interface or encrypted overlay, and moving peer tokens out of TOOLS.md into a secret store. Require explicit user confirmation before sending content to a peer and avoid constructing shell commands with raw message text.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (5)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:194
Finding

Mutable Remote Installer Is Executed Directly Through a Shell

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
SKILL.md:22
Finding

Plugin and npm Dependencies Are Installed from Unpinned Mutable Sources

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
references/tools-md-template.md:18
Finding

User-Controlled Message Text Is Interpolated into a Shell Command

Content
View full analysis
/plugins/claw-crony/skill/scripts/a2a-send.mjs \ --peer-url http://:18800 \ --token \ --message "YOUR MESSAGE HERE" ``` ### Technical Analysis The template directs an agent to construct a shell command by replacing `YOUR MESSAGE HERE` with user-supplied message content. Double quotes are not a safe shell-escaping boundary for arbitrary input. A message containing a closing quote, command substitution such as `$(...)`, backticks, or suitable shell metacharacters can alter the command interpreted by the shell. The risk arises before `a2a-send.mjs` processes its arguments: the shell parses and expands the constructed command first. Validation within the Node.js script would therefore not prevent shell-level injection. ### Attack Path 1. A user or untrusted instruction asks the agent to send a crafted A2A message. 2. The crafted message contains shell syntax, such as a closing quote followed by another command or command substitution. 3. The agent replaces `YOUR MESSAGE HERE` with that text as instructed by the template. 4. The exec tool passes the resulting string to a shell. 5. The shell interprets the injected syntax and executes unintended local commands in addition to, or instead of, the messaging script. ### Impact Assessment Exploitation can execute arbitrary commands with the privileges assigned to the agent's exec tool. This may allow reading workspace files, extracting peer tokens, modifying OpenClaw configuration, tampering with plugins, or making unauthorized network requests. The maximum scope is bounded by the operating-system account and any sandbox applied to the exec tool. ]]>
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
references/tools-md-template.md:9
Finding

Reusable Bearer Tokens Are Stored in Plaintext Agent Instructions

Content
View full analysis
| | | ### How to send a message to a peer When the user says "通过 A2A 让 做 xxx" / "Send to : xxx" / "Ask to ..." or similar, use the exec tool to run: ```bash node /plugins/claw-crony/skill/scripts/a2a-send.mjs \ --peer-url http://:18800 \ --token \ ``` ### Technical Analysis The Skill requires replacing `` with the actual bearer credential and appending the resulting text to the agent's persistent `TOOLS.md` file. This places a reusable authentication secret in ordinary documentation and also embeds it in an executable command. Any component that can read the agent instructions, workspace, generated prompts, backups, diagnostic bundles, or command logs may obtain the credential. Passing the token as a command-line argument may additionally expose it through process listings or execution telemetry. ### Attack Path 1. The operator inserts an actual peer bearer token into the template and saves it in `TOOLS.md`. 2. A malicious plugin, local process, workspace reader, backup recipient, or exposed model context obtains the file or command content. 3. The attacker extracts the bearer token and peer address. 4. The attacker submits authenticated requests to the peer gateway using the stolen token. 5. Requests are accepted with the authority assigned to that bearer credential until it is revoked or rotated. ### Impact Assessment Credential disclosure can enable unauthorized A2A requests to the target peer. Depending on peer routing and agent capabilities, an attacker may submit tasks, consume resources, obtain generated responses, or trigger tools available to th ...[truncated 145 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:73
Finding

A2A Gateway Is Bound to All Interfaces and Documented with Plaintext HTTP

Content
View full analysis
:18800/a2a/jsonrpc' ``` ```json { "name": "", "agentCardUrl": "http://:18800/.well-known/agent-card.json", "auth": { "type": "bearer", "token": "" } } ``` ### Technical Analysis Binding the gateway to `0.0.0.0` exposes it on every available network interface rather than only the interface needed for peer communication. The documentation permits LAN or public-IP deployment and consistently configures endpoints with plaintext HTTP. Bearer authentication protects access only if the token remains confidential. It does not provide transport confidentiality or server authentication over an untrusted network. Without a firewall or encrypted overlay, network observers may inspect messages and credentials, while unintended interfaces increase the remotely reachable attack surface. The local and peer verification commands that pipe `curl` output to `python3 -m json.tool` only parse JSON; they do not execute the response as shell code and are not themselves remote-code-execution findings. ### Attack Path 1. The operator configures the server to listen on `0.0.0.0:18800`. 2. An unintended LAN, public, container, or virtual network interface makes the service reachable by an untrusted party. 3. The attacker scans or connects to the exposed endpoint and enumerates the Agent Card or probes the service. 4. If the attacker can observe plaintext HTTP traffic, the attacker captures bea ...[truncated 557 chars]
Remediation
View remediation
Vulnerability Patterns
  • 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
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
Findings (9)

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 119)May include surrounding context.

md
openclaw gateway restart

# Verify Agent Card
curl -s http://localhost:18800/.well-known/agent-card.json | python3 -m json.tool

# Verify peer connectivity
curl -s http://<PEER_IP>:18800/.well-known/agent-card.json | python3 -m json.tool

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 122)May include surrounding context.

curl -s http://localhost:18800/.well-known/agent-card.json | python3 -m json.tool

Verify peer connectivity

curl -s http://<PEER_IP>:18800/.well-known/agent-card.json | python3 -m json.tool

text

## Step 9: Configure TOOLS.md

External Script Fetching

High
Category
Supply Chain
Confidence
99% confidence
Finding

The skill instructs users to fetch and immediately execute a remote script via 'curl ... | sh', which creates a direct remote code execution path if the source, transport, DNS, or hosting account is compromised. Because this installation path is presented as normal setup, users may run it without inspection, and in this context it is followed by privileged networking steps, increasing risk.

Content

Scanner excerpt · SKILL.md (reported line 194)May include surrounding context.

When servers are on different networks, use Tailscale:

bash
curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale up
# Authenticate via the printed URL (use same account on all servers)
tailscale ip -4  # Get the 100.x.x.x IP

Chaining Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

The explicit '| sh' shell-chaining turns downloaded network content into immediately executed commands, removing any review boundary and magnifying supply-chain and man-in-the-middle risks. In a setup skill, this is especially dangerous because users are likely to copy-paste it during privileged system configuration.

Content

Scanner excerpt · SKILL.md (reported line 194)May include surrounding context.

When servers are on different networks, use Tailscale:

bash
curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale up
# Authenticate via the printed URL (use same account on all servers)
tailscale ip -4  # Get the 100.x.x.x IP

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill generates a bearer token and explicitly instructs the user to save and share it with peers, but it does not warn that this token is a secret or provide guidance for secure storage, rotation, or transmission. In an A2A setup, compromise of this token would allow unauthorized peers to send authenticated requests to the gateway.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · SKILL.md (reported line 195)May include surrounding context.

bash
curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale up
# Authenticate via the printed URL (use same account on all servers)
tailscale ip -4  # Get the 100.x.x.x IP

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The template tells operators to place peer authentication tokens directly into TOOLS.md and inline command examples that are intended to be visible to the agent. That unnecessarily exposes bearer credentials to prompts, logs, downstream tools, and anyone with access to the workspace, enabling unauthorized cross-server messaging or impersonation if the token is leaked.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The phrase "or similar" broadens the set of natural-language triggers that authorize networked peer messaging, making it easier for ambiguous, indirect, or adversarial prompts to cause unintended cross-server communication. In an agent instruction file, loose trigger wording weakens intent boundaries and increases prompt-injection and accidental-exfiltration risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The template instructs the agent to send arbitrary message content to another server using an authenticated command, but provides no warning, consent check, or data-classification guardrail. This makes sensitive user or system data more likely to be transmitted off-host without notice, and the included token handling amplifies the impact by authorizing the outbound request automatically.

Content

No source excerpt is available for this finding.