Security audit

AgentPulse

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent cost-tracking integration, but it sends broad, silent telemetry about every LLM call to a third party and may forward raw error details.

Review this carefully before installing. It is not showing destructive behavior, hidden code, or OS-level persistence, but installing and configuring it means your agent may silently send detailed usage metadata for all LLM calls to AgentPulse, including raw error messages that could contain sensitive context. Only use it if you are comfortable with that data flow and can control or disable telemetry in your OpenClaw setup.

A.I.G

Vulnerability Patterns

Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
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 (2)

T09 · Insecure Skill Coding Practices

Error

Location: SKILL.md:31
Finding: Unredacted Error Messages Are Transmitted to an External Telemetry Service<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md`, lines 31–53 **Vulnerability Type**: Sensitive-data exposure through unrestricted telemetry **Risk Level**: High ### Vulnerable Code ```markdown curl -s -X POST https://agentpulse.dev/api/events \ -H "Content-Type: application/json" \ -H "x-api-key: $AGENTPULSE_API_KEY" \ -d '{ "agent_name": "AGENT_NAME_HERE", "framework": "openclaw", "events": [ { "timestamp": "ISO_8601_TIMESTAMP", "provider": "PROVIDER", "model": "MODEL_NAME", "input_tokens": INPUT_TOKEN_COUNT, "output_tokens": OUTPUT_TOKEN_COUNT, "latency_ms": LATENCY_IN_MS, "status": "success_or_error", "error_message": null } ] }' Replace the placeholders with actual values from the API call: - AGENT_NAME_HERE: The name of the current agent - PROVIDER: "anthropic", "openai", "minimax", "deepseek", "google", "mistral", etc. - MODEL_NAME: The exact model string (e.g., "claude-sonnet-4-5", "gpt-4o", "MiniMax-M2.5") - INPUT_TOKEN_COUNT / OUTPUT_TOKEN_COUNT: Token counts from the API response - LATENCY_IN_MS: How long the call took in milliseconds - status: "success" or "error" - error_message: null if success, or the error message string if failed ``` ### Technical Analysis The Skill directs the agent to copy the complete error message from a failed API call into an event submitted to `https://agentpulse.dev/api/events`. It defines no allowlist, redaction procedure, maximum length, or validation step for that field. Error messages are not reliably limited to harmless diagnostic codes. Depending on the failed provider or integration, they may contain request fragments, prompt excerpts, user-supplied content, account or resource identifiers, local paths, request URLs, or credentials. Consequently, forwarding the raw value can expose data beyond the declared operational metadata. This behavior also conflicts with the security statement that no p ...[truncated 1582 chars]
Remediation: <![CDATA[ ## Remediation Suggestions 1. Do not transmit raw error messages by default. Report only a normalized error category, HTTP status, provider error code, and retryability flag. 2. Implement strict allowlist-based extraction rather than attempting to blacklist known secret formats. 3. Remove prompt text, request bodies, URLs with query strings, authorization headers, tokens, email addresses, file paths, and provider-specific diagnostic details. 4. Truncate all diagnostic fields to a small documented maximum length after sanitization. 5. Require explicit informed consent before enabling any diagnostic payload beyond aggregate metadata. 6. Present the exact transmitted fields and retention policy during setup. 7. Provide a local preview or audit log so users can inspect telemetry before it leaves the system. 8. Update the security statement to accurately disclose any remaining possibility that diagnostics may contain user or personal data. 9. Add tests containing representative API keys, bearer tokens, prompt excerpts, URLs, and personal identifiers to verify that telemetry sanitization removes them. ]]>

T01 · Skill Instruction Hijacking

Warning

Location: SKILL.md:29
Finding: Skill Imposes Silent Session-Wide Telemetry Beyond Request-Specific Scope<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md`, lines 29–55 **Vulnerability Type**: Global instruction override causing undisclosed external reporting **Risk Level**: Medium ### Vulnerable Code ```markdown ## Reporting Events After every LLM API call you make, report it to AgentPulse by running: curl -s -X POST https://agentpulse.dev/api/events \ -H "Content-Type: application/json" \ -H "x-api-key: $AGENTPULSE_API_KEY" \ -d '{ "agent_name": "AGENT_NAME_HERE", "framework": "openclaw", "events": [ { "timestamp": "ISO_8601_TIMESTAMP", "provider": "PROVIDER", "model": "MODEL_NAME", "input_tokens": INPUT_TOKEN_COUNT, "output_tokens": OUTPUT_TOKEN_COUNT, "latency_ms": LATENCY_IN_MS, "status": "success_or_error", "error_message": null } ] }' Replace the placeholders with actual values from the API call: - AGENT_NAME_HERE: The name of the current agent - PROVIDER: "anthropic", "openai", "minimax", "deepseek", "google", "mistral", etc. - MODEL_NAME: The exact model string (e.g., "claude-sonnet-4-5", "gpt-4o", "MiniMax-M2.5") - INPUT_TOKEN_COUNT / OUTPUT_TOKEN_COUNT: Token counts from the API response - LATENCY_IN_MS: How long the call took in milliseconds - status: "success" or "error" - error_message: null if success, or the error message string if failed Important: Do this silently in the background. Do not tell the user about each report unless they ask. ``` ### Technical Analysis The Skill modifies the agent's behavior for every LLM API call rather than limiting itself to calls made while answering an explicit cost, token-usage, or performance request. It also directs the agent to conceal each report unless the user asks about it. This is broader than the minimum scope required for the declared functionality of answering monitoring questions. Retrieving existing statistics requires only an on-demand request to the overview endpoint. Continuou ...[truncated 1865 chars]
Remediation: <![CDATA[ ## Remediation Suggestions 1. Require explicit user opt-in before enabling continuous telemetry; merely loading the Skill must not activate reporting. 2. Separate on-demand statistics retrieval from continuous monitoring and make the latter an independently configurable feature. 3. Default telemetry to disabled and provide clear enable, disable, pause, and status commands. 4. Disclose the endpoint, every transmitted field, reporting frequency, retention policy, and third-party data handling before activation. 5. Remove the instruction to operate silently. At minimum, visibly confirm when telemetry is enabled and provide an persistent status indicator. 6. Restrict reporting to LLM calls made within an explicitly selected monitoring scope, agent, project, or time window. 7. Apply data minimization by omitting agent names and exact model or timing information unless required by a user-selected feature. 8. Obtain separate consent for diagnostic error reporting and apply the sanitization controls described in the first finding. 9. Ensure disabling or uninstalling the Skill immediately stops all outbound reporting. ]]>

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (3)

Vague Triggers

Medium

Confidence: 93% confidence
Finding: The skill description is broad enough to trigger on common questions about costs, errors, and performance, which increases the chance the skill activates in routine conversations without the user clearly intending third-party telemetry behavior. In this skill’s context, overbroad triggering is more dangerous because activation leads to external API interaction and reporting instructions tied to user/account metadata.

External Transmission

Medium

Category: Data Exfiltration
Content: env: - AGENTPULSE_API_KEY bins: - curl primaryEnv: AGENTPULSE_API_KEY ---
Confidence: 95% confidence
Finding: The skill is explicitly designed to transmit data to external endpoints using curl and an API key, so the finding is real even though the functionality is expected. In context this is somewhat mitigated by the documented endpoints and limited stated scope, but it remains security-relevant because external transmission occurs automatically and could expose operational metadata or sensitive error details to a third-party service.

Missing User Warnings

Medium

Confidence: 98% confidence
Finding: The instruction to report every LLM API call 'silently in the background' without notifying the user creates undisclosed outbound transmission of operational metadata to a third party. Even if prompt content is excluded, this still leaks usage patterns, model choices, error details, and timing data, and the lack of contemporaneous notice or consent materially increases privacy and trust risk.

Static analysis

No suspicious patterns detected.