Back to skill

Security audit

Ryder-Super-Pack

Security checks for vulnerabilities and agentic risk

Overview

This markdown-only skill is not clearly malicious, but it directs agents into broad high-impact workflows involving sensitive data, persistent memory, command execution, and external report delivery without enough user control.

Install only if you are comfortable with an agent using this skill for broad professional automation. Before use, require explicit confirmation for command execution, external searches, file writes, memory storage, archive moves, and any Telegram/Discord delivery; do not let it process finance, customer, legal, or personal data unless paths, recipients, retention, and review steps are clearly controlled.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (4)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
references/dev.md:16
Finding
Unrestricted Environment Variable Collection During Debugging<![CDATA[ ## Vulnerability Details **File Location**: `references/dev.md`, lines 16-20 **Vulnerability Type**: Excessive access to environment variables and secrets **Risk Level**: High ### Vulnerable Code Snippet ```markdown ## 3. Systematic Debugging (4-Phase) 1. **Root Cause**: Use `exec` to gather logs, environment variables, and trace data flow. 2. **Pattern**: Find working vs. broken code comparisons. 3. **Hypothesis**: Formulate "If I change X, then Y should happen." 4. **Fix & Verify**: Apply the fix and run the full test suite. ``` ### Technical Analysis The debugging workflow instructs the agent to use command execution to gather environment variables without limiting access to specific variables relevant to the fault under investigation. Process environments commonly contain API tokens, database connection strings, cloud credentials, webhook secrets, signing keys, and other authentication material. Reading the entire environment violates least-privilege principles because most debugging tasks require only a small, explicitly identified subset of configuration. Once collected, secret values may be exposed in command output, model context, session logs, generated debugging artifacts, or subagent context. The risk is amplified by other workflows in the Skill that support persistent memory and external report delivery. ### Attack Path 1. A debugging request causes the development reference to be loaded. 2. The agent follows the root-cause workflow and invokes `exec` to enumerate environment variables. 3. Credentials and other secrets become part of the tool output and agent context. 4. The values may subsequently be copied into diagnostics, logs, persistent memory, subagent prompts, or externally delivered reports. 5. Any party able to access those outputs could use the exposed credentials within the permissions granted to them. ### Impact Assessment This behavior may disclose every credential available to the agent process. The resulting acce ...[truncated 363 chars]
Remediation
<![CDATA[ ## Remediation Suggestions - Replace broad environment enumeration with an allowlist of explicitly named, task-relevant variables. - Never print the values of variables whose names indicate credentials, including names containing `TOKEN`, `SECRET`, `KEY`, `PASSWORD`, `COOKIE`, or `AUTH`. - Report only whether sensitive variables are present and, where necessary, their type or redacted metadata. - Require explicit user approval before inspecting potentially sensitive configuration. - Apply automatic secret redaction to command output, generated diagnostics, session logs, and subagent prompts. - Prevent environment-variable output from being written to persistent memory or transmitted through external channels. - Run diagnostic commands in a restricted environment containing only the minimum variables required by the target application. ]]>

T09 · Insecure Skill Coding Practices

Error
Location
references/finance.md:47
Finding
Sensitive Financial and Customer Data May Be Stored in Long-Term Agent Memory<![CDATA[ ## Vulnerability Details **File Location**: `references/finance.md`, lines 47-51; related persistence instructions in `references/ai-agent.md`, lines 35-40 **Vulnerability Type**: Uncontrolled persistence of sensitive financial and customer data **Risk Level**: High ### Vulnerable Code Snippet ```markdown ## 5. Data Engineering for Finance ### Financial Data Pipelines - **Automated Extraction**: Use `exec` to run scripts that fetch data from APIs (Stripe, HubSpot, etc.). - **Transformation**: Use the workspace as a staging area for cleaning and formatting data before loading into `memory/`. ``` The referenced memory mechanism is defined as follows: ```markdown ## 3. Tool-Integrated RAG OpenClaw's memory system is your RAG. - **Ingestion**: Use `write` to save high-value context to `memory/`. - **Retrieval**: Use `memory_search` with specific queries before every complex decision. - **Update**: Use `memory_get` to pull exactly what you need, avoiding context bloat. ``` The same finance reference also authorizes ingestion of bank statements, general-ledger extracts, transaction lists, ERP exports, and data obtained from Stripe or HubSpot. ### Technical Analysis The workflow moves transformed finance and customer information into a long-term memory directory without specifying data minimization, encryption, retention, deletion, access control, tenant isolation, or user approval. Stripe, HubSpot, bank-statement, ERP, and general-ledger data may contain personal identifiers, transaction histories, account information, customer communications, and confidential business metrics. Transformation does not inherently anonymize this information. This is not classified as agent memory poisoning because the instructions do not write attacker-controlled behavioral rules into memory. The confirmed issue is insecure persistence of sensitive data. ### Attack Path 1. The agent reads a bank statement, ERP export, transaction list, or customer dataset, or fe ...[truncated 997 chars]
Remediation
<![CDATA[ ## Remediation Suggestions - Default to ephemeral processing for financial, payment, CRM, and personal data. - Prohibit storing raw bank statements, transaction records, customer identifiers, credentials, or complete API responses in long-term memory. - Require explicit, informed user approval before persisting any derived information. - Persist only minimized and anonymized summaries that are necessary for a documented future purpose. - Encrypt approved persistent data at rest and enforce workspace, tenant, user, and subagent access boundaries. - Define retention periods and implement verified deletion after the task or retention period ends. - Record data provenance, purpose, classification, and expiration alongside every approved memory entry. - Add automated detection and redaction for account numbers, email addresses, payment identifiers, access tokens, and other sensitive fields. ]]>

T05 · Unauthorized Access and Privilege Escalation

Error
Location
references/legal.md:22
Finding
Data Subject Request Workflow Packages Personal Data Without Identity or Authority Verification<![CDATA[ ## Vulnerability Details **File Location**: `references/legal.md`, lines 22-28 **Vulnerability Type**: Missing authorization controls in personal-data disclosure workflow **Risk Level**: High ### Vulnerable Code Snippet ```markdown ### Automated DSR (Data Subject Request) Handling - **Intake**: Monitor for DSRs via `heartbeat` or manual trigger. - **Verification**: Use `web_search` to verify regulatory deadlines (e.g., 72 hours for GDPR breach). - **Fulfillment**: 1. **Search**: Use `grep` or `memory_search` to find all personal data related to the requester in the workspace. 2. **Package**: Use `write` to create a `DSR_RESPONSE_[ID].md`. 3. **Log**: Update a `COMPLIANCE_LOG.md` with the request status. ``` ### Technical Analysis Although the workflow includes a step named “Verification,” that step verifies regulatory deadlines rather than the requester's identity, legal authority, or entitlement to receive the requested data. The workflow then performs a broad search across both the workspace and persistent memory and consolidates matching personal information into a response file. This aggregation can increase exposure because information previously distributed across multiple sources becomes available in one artifact. The instruction does not define search boundaries, requester matching criteria, authorization checks, third-party data filtering, response-file permissions, or mandatory human review before disclosure. ### Attack Path 1. An attacker or unauthorized person submits a data subject request using another person's identity or an ambiguous identifier. 2. The request enters the workflow through a manual trigger or monitored intake. 3. The agent verifies only the applicable deadline and does not establish identity or authority. 4. The agent searches the workspace and long-term memory for all matching personal data. 5. Matching records are consolidated into `DSR_RESPONSE_[ID].md`. 6. The response artifact is accidentally or inte ...[truncated 537 chars]
Remediation
<![CDATA[ ## Remediation Suggestions - Add mandatory requester identity verification before any personal-data search begins. - Verify the authority of representatives, guardians, employees, or legal counsel acting for another person. - Require human approval after identity verification and before data collection or release. - Restrict searches to repositories and record types covered by the request and requester authorization. - Use exact, validated identifiers rather than broad name-based searches where possible. - Separate third-party information and legally exempt records from the response. - Store generated response packages in a restricted location with encryption and narrowly scoped permissions. - Add a final legal and privacy review before transmission. - Maintain an audit trail recording verification, search scope, reviewer approval, release destination, and deletion date. ]]>

T09 · Insecure Skill Coding Practices

Error
Location
references/marketing.md:32
Finding
Marketing Analytics Reports Are Automatically Sent to External Messaging Channels Without Review<![CDATA[ ## Vulnerability Details **File Location**: `references/marketing.md`, lines 32-37 **Vulnerability Type**: Uncontrolled external transmission of potentially sensitive analytics data **Risk Level**: High ### Vulnerable Code Snippet ```markdown ## 3. Analytics & Performance (Data-Driven Decisions) ### Tool-Integrated Analytics - **Data Ingestion**: Use `read` to ingest CSV exports of marketing data. - **Analysis**: Use `exec` with Python scripts to run cohort analysis or MMM (Marketing Mix Modeling). - **Reporting**: Automatically generate a `MARKETING_REPORT.md` and deliver it via the active channel (Telegram/Discord). ``` ### Technical Analysis The workflow ingests local marketing CSV files and automatically delivers a derived report through Telegram or Discord. It does not require a preview, user confirmation, data-classification check, destination validation, redaction, or recipient allowlist. Marketing exports can contain customer identifiers, cohort details, campaign costs, conversion information, revenue attribution, internal budgets, and confidential performance metrics. Generated reports can also accidentally contain source rows, file paths, command output, or credentials included in the input. The use of the “active channel” creates an additional destination-confusion risk because the active conversation may not be an approved recipient for the underlying dataset. ### Attack Path 1. The agent reads a marketing analytics CSV containing confidential metrics or customer-level data. 2. Python analysis generates cohort or marketing-mix results. 3. `MARKETING_REPORT.md` incorporates sensitive source fields or derived business information. 4. The workflow automatically delivers the report through the currently active Telegram or Discord channel. 5. If the active channel is shared, misconfigured, compromised, or not approved for the dataset, unauthorized recipients receive the report. ### Impact Assessment The external disclosure scope can ...[truncated 405 chars]
Remediation
<![CDATA[ ## Remediation Suggestions - Remove automatic delivery and save generated reports locally by default. - Present the complete report and destination to the user before transmission. - Require explicit confirmation for every external send involving local or derived data. - Validate that the destination is approved for the relevant data classification and organization. - Use an allowlist of authorized channel identifiers rather than the implicitly active channel. - Remove customer-level records and redact identifiers, credentials, file paths, and confidential source fields. - Apply data-loss-prevention checks before invoking Telegram, Discord, or any other external messaging integration. - Log the approved recipient, approving user, report hash, transmission time, and applicable retention policy. ]]>
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
Findings (27)

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Content
---
name: ryder-super-pack
description: Optimized super-skill collection for OpenClaw/Codex, merging Perplexity + Claude Code expertise across 11 domains. Features specialized reference loading (progressive disclosure), tool-integrated workflows (exec, fs, web_fetch), and persona-aligned frameworks. Use for high-stakes professional AI tasks requiring structured reasoning, templates, or domain-specific gap analysis.
---

# Ryder Super-Pack (汪汪队超级包) - OpenClaw Edition

Greetings, Mayor. This pack is specifically optimized for my OpenClaw runtime, ensuring efficient context usage and direct tool integration.

#
Confidence
80% confidence
Finding
YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Vague Triggers

Medium
Confidence
93% confidence
Finding
The skill description is extremely broad and markets itself for 'high-stakes professional AI tasks' across many domains, which increases the chance it will be invoked for ordinary or sensitive requests without clear scope boundaries. In an agent ecosystem, overbroad activation criteria can cause unnecessary tool access and unsafe delegation into workflows involving exec, fs, or web capabilities.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The markdown advertises tool-first execution using exec, fs, web_fetch/web_search, and subagent orchestration, but it does not include any safety constraints, approval requirements, or warnings about modifying files, running commands, or pulling untrusted remote content. This creates a realistic path for unsafe execution or data exposure if the skill is triggered in a permissive runtime.

Missing User Warnings

Medium
Confidence
90% confidence
Finding
The skill directs the agent to generate a `VARIANCE_REPORT.md` file without warning the user that new artifacts will be created in the workspace. In high-stakes finance contexts, silent file creation can mislead users about what was produced, overwrite expected outputs, or leave sensitive financial analysis in shared workspaces without informed consent.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The workflow instructs automatic generation of `JOURNAL_ENTRIES.md` and maintenance of `CLOSE_STATUS.md` without warning or confirmation. Because these files can influence accounting operations and internal close processes, silent creation may cause users to rely on unreviewed outputs or expose sensitive operational status information in the workspace.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The market-data workflow directs use of `web_search` and `web_fetch` against external financial sources without warning that the agent will initiate network access. In investment-research settings, undisclosed outbound requests can leak the user’s research interests, trigger compliance concerns, or fetch untrusted content that is then treated as analysis input.

Context-Inappropriate Capability

Medium
Confidence
86% confidence
Finding
The manifest describes a structured super-skill collection for domain-specific reasoning and reference loading, and this file is specifically a finance reference. In that context, instructing the agent to use `exec` to run scripts that fetch data from arbitrary APIs such as Stripe and HubSpot expands from financial analysis guidance into generalized integration and data-engineering automation, which is not obviously required by the stated finance-reference purpose.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The data-engineering section instructs running scripts to fetch data from third-party APIs and stage it in the workspace or `memory/` without any disclosure about external access or data movement. In a finance context this is more dangerous because it may involve sensitive customer, billing, or CRM data, creating risk of unintended transfer, retention, or policy violations if done without explicit authorization.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
This workflow directs the agent to search for all personal data related to a requester and then write response and compliance log files, but it provides no safeguards around minimization, access control, verification strength, or safe handling of sensitive data. In an agent environment with filesystem and search tools, this can lead to over-collection, unnecessary duplication of personal data, and accidental disclosure into workspace artifacts.

Missing User Warnings

Medium
Confidence
91% confidence
Finding
The skill instructs automatic delivery of generated marketing reports via Telegram/Discord, which can transmit analytics, campaign data, or other sensitive business information to external platforms without explicit user consent or a sensitivity check. In this context, the skill is operational and tool-enabled, so the omission is more dangerous because it normalizes exfiltration of potentially confidential data as part of routine workflow.

Context-Inappropriate Capability

Medium
Confidence
92% confidence
Finding
This reference instructs the agent to use exec-driven automation for portfolio prioritization and to move stale files into an archive, which crosses from passive guidance into autonomous system and filesystem modification. In a general-purpose super-skill, these actions can trigger unintended subprocess execution or data relocation without clear user consent, scope limits, or safety checks.

Missing User Warnings

Medium
Confidence
91% confidence
Finding
The cleanup instruction tells the agent to move stale files to an archive based on age alone, but does not require warning the user or presenting a review list before relocation. This can cause operational disruption, hidden data movement, or loss of discoverability if important files are archived incorrectly.

Context-Inappropriate Capability

Medium
Confidence
94% confidence
Finding
The invoice workflow directs the agent to read invoice files and persist extracted financial data into a tracker, creating a pathway for handling sensitive financial information without privacy controls or authorization checks. In a broad super-skill context, this is risky because it normalizes persistent processing of potentially confidential records as a default operational behavior.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The invoice extraction workflow processes potentially sensitive billing documents and writes derived data to a finance tracker without any privacy warning, data-classification note, or consent requirement. This increases the likelihood of unauthorized exposure, over-collection, or accidental persistence of confidential financial information.

Missing User Warnings

Medium
Confidence
96% confidence
Finding
The skill tells the agent to ingest interview transcripts and survey results and derive personas from them without any privacy or data-handling warning. User research data often contains sensitive personal, behavioral, or confidential business information, so processing it without caution increases the risk of over-collection, retention, or exposure in generated artifacts.

Context-Inappropriate Capability

Medium
Confidence
83% confidence
Finding
The manifest frames the skill as a professional super-skill collection for structured reasoning and domain workflows, but this file specifically instructs use of `exec` for Python scripts, `pdftotext`, and `yt-dlp`. While tool integration is mentioned generally, encouraging arbitrary command execution for media extraction and local processing is a materially stronger capability than the stated research-and-synthesis purpose alone justifies.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The skill tells the agent to check `memory/` for prior interactions without warning the user that historical interaction data may be accessed. In a sales context, that data may contain sensitive notes, personal information, or prior customer communications, making undisclosed access a privacy and confidentiality risk.

Missing User Warnings

Low
Confidence
84% confidence
Finding
The skill directs the agent to use web search and read tools to compile background summaries on external participants, which can affect individual privacy and may involve collecting personal information. The markdown does not include any caution about limiting collection, verifying relevance, or complying with privacy expectations.

Description-Behavior Mismatch

Low
Confidence
79% confidence
Finding
The manifest describes a structured, domain-specific skill collection for high-stakes professional tasks, and this file is framed as a legal operations reference. However, the instruction to use web search to summarize the background of all external participants expands from legal operations into general people/background research, which is not clearly part of the stated legal purpose of this reference.

Natural-Language Policy Violations

Low
Confidence
1% confidence
Finding
After review, no genuine language or locale policy violation is present in this file. The content does not force a specific human language or locale, so no SQP-3 finding should be reported.

Missing User Warnings

Low
Confidence
86% confidence
Finding
L12 directs creation of `content-brief.md`, and similar write behavior is also described later for `MARKETING_REPORT.md` and `BATTLECARD.md`. The skill description does not warn users that it will create files in the workspace, which is a user-data and system-state affecting behavior that should be disclosed in markdown.

Missing User Warnings

Low
Confidence
94% confidence
Finding
The workflow directs use of `web_search` to gather competitor features and user pain points, but does not disclose that this involves external network access. In a high-stakes professional skill, undisclosed outbound access can violate user expectations, leak contextual information through queries, or breach restricted-environment policies.

Missing User Warnings

Low
Confidence
90% confidence
Finding
The skill explicitly instructs the agent to create and update workspace files like `PRD_[FEATURE].md` without telling the user that local files may be modified. In an agent setting, silent file writes can surprise users, overwrite existing work, or create unintended artifacts, even if the action is framed as normal PM workflow.

Description-Behavior Mismatch

Low
Confidence
78% confidence
Finding
The manifest describes a collection for structured reasoning, templates, and domain-specific analysis, while this reference goes beyond pure research guidance by directing creation of concrete files such as `RESEARCH_BRIEF.md` and storage in `memory/` for long-term retrieval. Those write-oriented operations are not highlighted in the manifest's description of a research/knowledge workflow.

Missing User Warnings

Low
Confidence
91% confidence
Finding
The skill instructs the agent to create a `RESEARCH_[COMPANY].md` file without disclosing that it will write to local storage or asking for user approval. Even though this is a normal sales workflow, undisclosed file creation can surprise users, overwrite existing files, or leave sensitive prospect research on disk.

Static analysis

No suspicious patterns detected.