Back to skill

Security audit

Clawguard Skill

Security checks for vulnerabilities and agentic risk

Overview

This DLP skill has a coherent security purpose, but it installs a persistent network-facing service and has defaults that can expose or fail to protect sensitive outbound content.

Review before installing. Use only if you are comfortable with a plugin that intercepts outbound tool calls and runs a long-lived local service. Prefer changing the bind address to 127.0.0.1, setting blockOnError to true, removing the unpinned PyPI fallback, and making systemd/launchd persistence an explicit opt-in.

Vulnerability Patterns
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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)

T06 · System Persistence

Error
Location
setup.sh:198
Finding

Redundant cross-session persistence through systemd and launchd

Content
View full analysis
/dev/null; then # ── systemd user service (Linux) ── SYSTEMD_DIR="$HOME/.config/systemd/user" mkdir -p "$SYSTEMD_DIR" cat > "$SYSTEMD_DIR/clawwall.service" </dev/null || true systemctl --user enable clawwall.service 2>/dev/null || true ok "Created systemd user service: $SYSTEMD_DIR/clawwall.service" elif [[ "$(uname)" == "Darwin" ]]; then # ── launchd plist (macOS) ── PLIST_DIR="$HOME/Library/LaunchAgents" PLIST_FILE="$PLIST_DIR/com.clawwall.service.plist" mkdir -p "$PLIST_DIR" cat > "$PLIST_FILE" < Label com.clawwall.service ProgramArguments $CLAWWALL_BIN RunAtLoad KeepAlive EnvironmentVariables CLAWGUARD_HOST 0.0.0.0 CLAWGUARD_PORT $CLAWWALL_PORT StandardOutPath $CLAWWALL_CONFIG_DIR/clawwall.log StandardErrorPath $CLAWWALL_CONFIG_DIR/clawwall.err PLISTEOF ``` ```bash if [[ "$(uname)" == "Linux" ]] && command -v systemctl &>/dev/null; then systemctl --user start clawwall.service 2>/dev/nu ...[truncated 1852 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
setup.sh:203
Finding

DLP service binds to all network interfaces despite local-only claims

Content
View full analysis
"$SYSTEMD_DIR/clawwall.service" <EnvironmentVariables CLAWGUARD_HOST 0.0.0.0 CLAWGUARD_PORT 8642 ``` ### Technical Analysis Both generated service definitions set `CLAWGUARD_HOST=0.0.0.0`, causing the service to listen on every available IPv4 interface if the installed Python service honors this variable. This exceeds the plugin's requirements because the plugin connects to `http://127.0.0.1:8642`. It also conflicts with the documentation's assertion that the service is “fully local.” Binding to all interfaces can make the scan API and dashboard reachable from other hosts on the same network or any other routed network not blocked by a host firewall. The Python service is not included in the audited artifact. Consequently, the presence of authentication, authorization, origin validation, and endpoint protections could not be verified. The exposure is confirmed at the generated configuration level, while the exact remotely exploitable operations depend on the external service implementation. ### Attack Path 1. The user runs the installer. 2. The generated service starts ClawWall with `CLAWGUARD_HOST=0.0.0.0`. 3. Port 8642 listens on non-loopback interfaces. 4. A host with network reachability connects to port 8642. 5. If the service lacks adequate authentication, the remote host can access the scan API or dashboard and interact with exposed service functionality. ### Impact Assessment Potential im ...[truncated 338 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
plugin/index.ts:165
Finding

Incorrect redaction mapping can allow sensitive arguments to leave unchanged

Content
View full analysis
): string { const parts: string[] = []; for (const [key, value] of Object.entries(args)) { if (typeof value === "string") { parts.push(value); } else if (value !== null && value !== undefined) { parts.push(JSON.stringify(value)); } } return parts.join("\n"); } ``` ```ts case "REDACT": if (scan.content !== null) { // Replace all string args with the redacted content // This is a simplified approach — a production version would // map redaction offsets back to individual args const modifiedArgs = { ...context.args }; for (const key of Object.keys(modifiedArgs)) { if (typeof modifiedArgs[key] === "string") { modifiedArgs[key] = scan.content; break; // Only replace the first string arg for now } } return { allow: true, modifiedArgs }; } return { allow: true }; ``` ```ts if (scan.suggested_action === "REDACT" && scan.content !== null) { const modifiedArgs = { ...context.args }; for (const key of Object.keys(modifiedArgs)) { if (typeof modifiedArgs[key] === "string") { modifiedArgs[key] = scan.content; break; } } return { allow: true, modifiedArgs }; } ``` ### Technical Analysis The plugin serializes every argument into one newline-delimited string and submits that combined string for scanning. When the service returns redacted content, the plugin does not map the redacted offsets back to their original arguments. Instead, it replaces only the first top-level string argument and immediately exits the loop. This produces two security failures: 1. A secret in a later string argument remains unchanged. 2. Secrets inside arrays or nested objects remain unchanged because only top-le ...[truncated 1095 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
plugin/index.ts:46
Finding

DLP enforcement fails open on service errors and unexpected responses

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
setup.sh:74
Finding

Unpinned fallback installs and persists unreviewed external code

Content
View full analysis
/dev/null; then ok "Installed clawwall $CLAWWALL_VERSION from PyPI" elif pip install clawwall 2>/dev/null; then ok "Installed clawwall from PyPI (latest)" else # Fallback: install from local source if we're in the repo REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)" if [ -f "$REPO_ROOT/pyproject.toml" ]; then info "PyPI install failed — installing from local source..." pip install "$REPO_ROOT" ok "Installed clawwall from local source" else fail "Could not install clawwall. Run 'pip install clawwall' manually." fi fi ``` ### Technical Analysis The core Python service is not present in the audited artifact. The installer retrieves it from PyPI at installation time. Although it first requests version `0.3.0`, any failure causes a silent fallback to `pip install clawwall`, which resolves the latest available version without a version pin or hash. This means the effective runtime code can differ from the reviewed skill. The service receives all content sent for DLP scanning and is subsequently configured for automatic startup, making its integrity particularly security-sensitive. The audit did not establish that the current PyPI package is malicious. The confirmed issue is that the installation process permits unreviewed, mutable third-party code to become the persisted security service. ### Attack Path 1. The pinned installation fails because version `0.3.0` is unavailable, incompatible, temporarily inaccessible, or otherwise fails to install. 2. The installer automatically requests the latest `clawwall` package. 3. PyPI supplies whatever version currently satisfies the unpinned request. 4. The installer creates persistent startup entries for the re ...[truncated 629 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (47)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill advertises a local privacy/security guard, yet also documents software installation from external package ecosystems, plugin/hook registration, modification of user config, autostart persistence, and a networked service that by default binds to 0.0.0.0. This broader operational footprint materially increases risk because compromise or misuse would gain persistence and expose a service interface beyond localhost, contrary to the narrow expectation set by the description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill advertises a local privacy/security guard, yet also documents software installation from external package ecosystems, plugin/hook registration, modification of user config, autostart persistence, and a networked service that by default binds to 0.0.0.0. This broader operational footprint materially increases risk because compromise or misuse would gain persistence and expose a service interface beyond localhost, contrary to the narrow expectation set by the description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill advertises a local privacy/security guard, yet also documents software installation from external package ecosystems, plugin/hook registration, modification of user config, autostart persistence, and a networked service that by default binds to 0.0.0.0. This broader operational footprint materially increases risk because compromise or misuse would gain persistence and expose a service interface beyond localhost, contrary to the narrow expectation set by the description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill advertises a local privacy/security guard, yet also documents software installation from external package ecosystems, plugin/hook registration, modification of user config, autostart persistence, and a networked service that by default binds to 0.0.0.0. This broader operational footprint materially increases risk because compromise or misuse would gain persistence and expose a service interface beyond localhost, contrary to the narrow expectation set by the description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill advertises a local privacy/security guard, yet also documents software installation from external package ecosystems, plugin/hook registration, modification of user config, autostart persistence, and a networked service that by default binds to 0.0.0.0. This broader operational footprint materially increases risk because compromise or misuse would gain persistence and expose a service interface beyond localhost, contrary to the narrow expectation set by the description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill advertises a local privacy/security guard, yet also documents software installation from external package ecosystems, plugin/hook registration, modification of user config, autostart persistence, and a networked service that by default binds to 0.0.0.0. This broader operational footprint materially increases risk because compromise or misuse would gain persistence and expose a service interface beyond localhost, contrary to the narrow expectation set by the description.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

When the scanner is unreachable, the plugin returns allow:true unless blockOnError is enabled, creating a fail-open path that can let secrets or PII leave the machine unscanned. In a DLP control whose stated purpose is preventing outbound leakage, availability failures become a bypass condition, which materially weakens the protection boundary.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The manifest advertises a local, no-LLM outbound DLP layer, but its configuration explicitly sends scanned content to a configurable service URL over plain HTTP. That creates a trust-boundary mismatch: sensitive outbound content may leave the machine and be exposed to interception, redirection, or unexpected remote processing, directly undermining the privacy guarantees implied by the skill description.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The installer configures the service to bind to 0.0.0.0 via both systemd and launchd, exposing the DLP API on all network interfaces instead of keeping it localhost-only. For a component described as local outbound blocking, this unnecessarily broadens the attack surface and could let other hosts access the service, dashboard, or health/API endpoints depending on host firewall posture.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
80% confidence
Finding

The skill declares broad capabilities through its documented behavior and startup hook, including shell execution, local networking, and environment-dependent installation, but does not declare any tool scope restrictions such as permissions or allowed-tools. This increases the attack surface because a user or platform cannot easily constrain what the skill may invoke during setup and startup, especially given that it installs software, edits configuration, and runs a persistent service.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 59)May include surrounding context.

bash
# Health check
curl -s http://127.0.0.1:8642/api/v1/health

# Test scan
curl -s -X POST http://127.0.0.1:8642/api/v1/scan \

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The code claims to ensure the ClawWall service is running before tool calls, but it only checks whether port 8642 is occupied and returns early if any process is listening. An attacker or unrelated local process can bind that port, causing the hook to falsely assume DLP protection is active, which could disable the intended outbound filtering control.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The hook executes which clawwall via a shell and then spawns whatever binary is found on the user's PATH as a detached background process. In a startup hook, this creates an execution surface broader than the stated regex-based DLP purpose: a PATH-hijacked or trojaned clawwall binary would be launched automatically and persist independently of the parent process.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The manifest describes a local outbound DLP tool with hard regex blocking and no LLM, which does not by itself justify a hook that discovers binaries via shell execution and launches a persistent detached service. This file adds process-management behavior that is operationally significant and not clearly implied by the skill's stated filtering purpose.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The code launches a detached background service with the full inherited environment and no explicit user-facing consent or warning at execution time. If sensitive variables are present in process.env, they become available to the child process, and the detached nature makes the subprocess harder for users to notice, audit, or stop.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The plugin sends all extracted outbound tool-call content to a separate local HTTP service for scanning, but this file provides no explicit disclosure, consent, or minimization before transmitting potentially sensitive data. Even though the endpoint is loopback-only, localhost services can still be monitored, replaced, or exposed by local compromise, port-forwarding, container boundary issues, or misconfiguration, making this a real privacy and data-handling risk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest describes the skill as "Outbound DLP ... hard regex blocks secrets & PII from leaving the machine. Domain control, no LLM," which implies deterministic local blocking behavior. In code, outbound content is sent to a separate service for adjudication, and the returned action can be ALLOW, REDACT, or PROMPT rather than a strict hard block, so the implemented behavior is broader and less absolute than described.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The plugin transmits all extracted outbound content to a scanning service over HTTP without any user-facing disclosure or consent on normal operation. Even though the default endpoint is localhost, this still expands the trust boundary and can expose sensitive data to another process or, if reconfigured, to a remote service, which is significant in a DLP context handling secrets and PII.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The REDACT path only replaces the first string argument, despite comments implying broader coverage, so sensitive data in additional string fields may still be sent unchanged. This creates a partial-redaction bypass where a tool call with multiple outbound string parameters can leak the very content the scanner flagged.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill description suggests hard outbound blocking, but the default configuration sets blockOnError to false, meaning scans fail open when the service is unreachable. An attacker or fault condition that disrupts the scanner can therefore bypass DLP enforcement and allow sensitive data to leave the system uninspected.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The package metadata says outbound tool calls are intercepted and scanned via a ClawWall service, which conflicts with the stated skill description of local-only, no-LLM DLP. That mismatch is security-relevant because users may rely on the manifest’s privacy boundary, while the implementation appears to send sensitive outbound content to an external service, expanding the trust boundary and creating possible data exfiltration risk.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
81% confidence
Finding

The script installs Python and Node packages from PyPI and npm as part of setup. While this may be operationally convenient for installation, fetching and executing third-party package code is not part of the core stated purpose of regex-based outbound DLP and introduces networked supply-chain capability beyond that purpose.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The installer silently modifies persistent user configuration and registers auto-start services without a dedicated consent step beyond generic setup output. This can surprise users, create stealthy persistence, and make rollback difficult, especially in an agent-skill ecosystem where users may expect narrower, temporary behavior.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

This portion of the script both configures network exposure on 0.0.0.0 and starts the service in the background automatically, without a strong safety warning or explicit opt-in. Combining silent persistence with network-accessible service startup materially increases risk because users may unknowingly deploy a reachable daemon on their machine.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
86% confidence
Finding

Enabling a user systemd service creates persistent execution across sessions. Persistence is not inherently malicious for legitimate software, but in this context it is security-relevant because it is installed automatically and runs a network service, increasing the chance of unnoticed long-term exposure.

Content

Scanner excerpt · setup.sh (reported line 221)May include surrounding context.

sh
SVCEOF

  systemctl --user daemon-reload 2>/dev/null || true
  systemctl --user enable clawwall.service 2>/dev/null || true
  ok "Created systemd user service: $SYSTEMD_DIR/clawwall.service"

elif [[ "$(uname)" == "Darwin" ]]; then