Back to skill

Security audit

My Research Lab

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent automated research assistant, but it asks for persistent automation, memory use, local scanning, storage, and outbound delivery with too much silent or under-scoped behavior.

Install only if you want an always-on research workflow. Before enabling it, explicitly choose the research scope, push destinations, archive location, schedule, and whether memory-derived personalization is allowed. Do not store SMTP_PASS or webhook secrets in lab-config.json, memory, reports, or archives; use a protected secret mechanism. Prefer pinned installer versions and review any recommended browser-control or source skills before adding them.

Vulnerability Patterns
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (2)

T08 · Insecure Dependencies

Warning
Location
README.md:39
Finding

Mutable and Unpinned Third-Party Installation Commands

Content
View full analysis

Vulnerability Details

File Location: README.md:39
Additional Location: references/channels-and-tools.md:72
Vulnerability Type: Supply-chain exposure through mutable dependency resolution
Risk Level: Medium

Vulnerable Code

README.md:38-40:

bash
npx clawhub@latest install my-research-lab

references/channels-and-tools.md:72:

bash
openclaw skills install web-access

Technical Analysis

The primary installation command explicitly resolves and executes the package tagged latest. This tag is mutable and can point to a different release after the reviewed Skill version has been published. The documented installation of web-access is likewise unversioned and does not specify a release digest, integrity hash, signed artifact, or immutable source revision.

As a result, auditing the current documentation or repository does not guarantee that the code eventually downloaded and executed by these commands is the same code that was reviewed. User confirmation for installing web-access reduces unauthorized installation risk, but it does not protect against a compromised publisher account, registry, future release, or transitive dependency.

The risk is elevated for web-access because the documentation describes it as controlling a real browser through Playwright and supporting authenticated sessions. A compromised version could potentially inherit access to browser data and active sessions available to the invoking environment.

Attack Path

  1. An attacker compromises the package publisher, registry account, release pipeline, or an upstream dependency.
  2. The attacker publishes a malicious package version and assigns it the latest tag, or replaces the unversioned web-access package release.
  3. A user follows the documented installation command.
  4. The package manager retrieves and executes the attacker-controlled release.
  5. Malicious installation or runtime log ...[truncated 931 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace @latest with an exact, reviewed package version.
  2. Pin web-access to an exact version or immutable source commit.
  3. Publish expected cryptographic hashes and verify downloaded artifacts before execution.
  4. Use signed releases and document how users can verify publisher signatures.
  5. Maintain a lockfile for any transitive dependencies used by executable components.
  6. Identify the authoritative repository and registry namespace for every recommended Skill.
  7. Require a fresh security review before upgrading pinned versions.
  8. Isolate browser-control Skills using a dedicated browser profile without unrelated authenticated sessions.
  9. Grant browser, filesystem, and network permissions only when required for a specific user-approved operation.

T09 · Insecure Skill Coding Practices

Warning
Location
references/channels-and-tools.md:14
Finding

Ambiguous Handling May Persist SMTP Authorization Secrets in Plaintext

Content
View full analysis

Vulnerability Details

File Location: references/channels-and-tools.md:14-16
Supporting Locations: references/config-schema.md:1, SKILL.md:8-14
Vulnerability Type: Potential plaintext credential persistence
Risk Level: Medium

Vulnerable Code

Faithful English rendering of references/channels-and-tools.md:14-16:

text
5. Enter the following in the research-lab configuration:
   - SMTP_USER: your QQ email address
   - SMTP_PASS: authorization code (not the login password)

SKILL.md:8-14 declares the same values as environment secrets while also granting access to persistent memory and automation storage:

yaml
requires:
  env:
    - SMTP_USER
    - SMTP_PASS
  config:
    - "~/.codebuddy/memory"
    - "~/.codebuddy/automations"

Faithful English rendering of references/config-schema.md:1:

text
Each research-lab instance maintains an independent lab-config.json in the memory system. A copy is also retained in the local archive directory.

Technical Analysis

SMTP_PASS is an SMTP authorization credential and should be treated as a secret. The Skill metadata appropriately identifies it as an environment variable, but the channel setup guide ambiguously instructs the user to enter it into the “research-lab configuration.”

Elsewhere, the project states that lab-config.json is stored in the memory system and copied to a local archive. It does not explicitly prohibit placing SMTP_PASS in that configuration, require a secret manager, define a secret-reference field, mandate restrictive file permissions, or require credential redaction from prompts and error logs.

This inconsistency can cause an implementation or Agent to persist the authorization code in memory, a JSON file, an archive, a backup, or a synchronized storage service. The schema example does not itself include SMTP_PASS, but the setup wording leaves the storage boundary insufficie ...[truncated 1345 chars]

Remediation
View remediation

Remediation Suggestions

  1. State explicitly that SMTP_PASS must only be stored through the platform secret manager or protected environment facility.
  2. Prohibit writing SMTP credentials to Agent memory, prompts, lab-config.json, archives, reports, logs, or cloud document storage.
  3. Store only an opaque secret reference, such as smtpSecretRef, in persistent configuration.
  4. Add schema validation that rejects fields named SMTP_PASS, password, authorizationCode, or equivalent secret-bearing values.
  5. Redact credentials from errors, diagnostic output, automation prompts, and tool-call traces.
  6. Apply least-privilege filesystem permissions to any secret storage used by the host platform.
  7. Avoid copying secret-bearing configuration into the local research archive.
  8. Document credential rotation and revocation procedures.
  9. Test channel availability without displaying or persisting the supplied authorization code.
  10. Separate ordinary research configuration from credential storage in both documentation and implementation.
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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (29)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README promises that after the user provides a research direction, the skill will '自动配置一切' and continuously collect, analyze, and produce output, including detecting push channels, storage platforms, and using user memory for association. Without prominent warnings and consent boundaries, this presents privacy and autonomy risks because the skill appears designed to access external channels, retained memory, and automation features by default.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The README instructs users to run npx clawhub@latest install my-research-lab, which fetches and executes the latest publisher-controlled package version at install time. This creates a supply-chain risk: a compromised or malicious future release of clawhub could execute arbitrary code on the user's machine, and the skill context encourages direct execution from documentation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The activation examples are broad and generic, such as '启动研究院' and topic-based phrases, making it unclear when the skill should trigger versus normal conversation. In an agent ecosystem, ambiguous triggers can cause unintended activation, leading the skill to begin automated collection, configuration, or task orchestration without sufficiently explicit user consent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger phrases are broad enough to match ordinary conversation such as general requests for updates, analysis, or brainstorming. In a skill that can read/write memory, create scheduled tasks, and perform network/local actions, unintended activation can cause the agent to begin persistent workflows or collect data without clear user intent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The permission section lists powerful capabilities, but the skill lacks a prominent upfront warning that it will automatically read/write memory, create scheduled jobs, access local files, and use network/push channels. Users may invoke the skill expecting a normal chat workflow, not realizing it can establish persistent automation and interact with sensitive local and external resources.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill instructs persistent storage and reuse of user memory, feedback history, and project progress as core operating data. Without strict minimization and scoping, this creates a natural-language leakage path where sensitive preferences, history, or contextual details may be surfaced in later summaries, analyses, or cross-project outputs.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The workflow explicitly tells the agent to generate summaries, add user-linked analysis, push them through email/IM, and archive them in external or local storage. This materially increases exfiltration risk because collected content and user-associated insights may be distributed across multiple channels, potentially exposing sensitive information beyond the original chat context.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The self-evolution design requires writing behavioral feedback, source-weighting decisions, and action outcomes back into memory after each run. Over time this builds a persistent profile of user behavior and preferences that can reveal sensitive patterns and can later influence outputs in ways the user may not expect or be able to audit.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The title and all instructional content are written entirely in Chinese, and the file does not indicate that language selection is optional or limited to a China-specific deployment context. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill instructs reuse of user memory and configuration and uses that state to steer setup decisions without clearly defining consent boundaries. This can operationalize previously stored personal context for a new automation flow in ways the user may not expect, especially when it influences channels, topics, or downstream actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill instructs silent auto-detection and reuse of configured push channels without an upfront notice or explicit opt-in. This can expose the existence of connected communication integrations and trigger outbound delivery over channels the user did not knowingly authorize for this workflow, which is a privacy and unintended-action risk.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Silent detection of messaging and storage channels can expose private configuration state and immediately make it actionable for delivery and persistence. The combination of hidden discovery plus operational use increases privacy risk because the user may not realize which integrations exist or are being activated.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill requires persistent storage of a broad set of materials, including full briefings, experiment records, project documents, commentary archives, and even distilled skills. This establishes expansive retention by default, increasing the chance that sensitive research content, internal reasoning, or proprietary material is stored longer and more broadly than necessary.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The storage setup performs silent backend detection and may default to local filesystem persistence without a clear warning. That creates undisclosed data-retention behavior and may place sensitive research outputs in locations the user did not expect or approve.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill directs silent filesystem scanning of local skill directories to discover and semantically match other installed skills. Even if limited to known paths, this inspects local environment state without notice and can reveal installed tools, interests, or configurations that the user did not intend to expose to this workflow.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill mandates creation and verification of scheduled tasks as a required startup step, but does not pair this with explicit user-facing warning or confirmation about modifying automation settings. Creating recurring jobs without clear consent can lead to persistent autonomous activity, repeated outbound messages, and ongoing storage of user data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The prompt instructs the agent to perform network collection across multiple sources and call domain-specific external skills, yet it provides no privacy notice, approval checkpoint, or restriction on what user context may be sent during those operations. In a research assistant that reads memory and personal configuration, this can expose sensitive preferences, interests, or derived data to external services during automated recurring execution.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The template directs the agent to automatically push content to configured channels, archive full reports, and write operational data into memory without requiring an explicit user-facing warning or per-action consent. This creates a real risk of unintended disclosure, persistence of sensitive data, and side effects on external systems, especially because the task is designed to run on a schedule without interactive confirmation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly instructs the agent to silently record weekly source-health summaries to memory without notifying the user. Silent persistence creates a privacy and transparency risk because users may not realize behavioral data and operational summaries are being retained across sessions.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instruction to 'silently' log weekly source-health information to memory establishes persistent natural-language logging without user awareness. Even if the content seems operational, it normalizes hidden retention behavior and can combine with other stored context to form a broader user or activity profile.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

These rules direct the agent to write feedback-derived rules and inferred preferences into memory after user interactions, again without clear user disclosure. Persisting inferred preferences and custom filtering logic can accumulate sensitive behavioral profiles and influence future outputs in ways the user did not knowingly authorize.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The feedback workflow requires persistent storage of detailed user feedback, inferred preferences, and generated filtering rules after each interaction. In a research-assistant skill, this is more dangerous because the retained data can reveal interests, priorities, dissatisfaction patterns, and ongoing project direction over time without explicit consent boundaries.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The fallback behavior stores failed push content as a local file but does not warn about local retention, file location sensitivity, or access controls. This can expose research notes, summaries, or other potentially sensitive content to unintended users or processes on the host system.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown file is entirely presented in Chinese, including operational guidance and user-facing prompt text, but it does not state that the skill is intended only for Chinese-speaking users or provide any language/locale opt-in. That can violate a language-policy requirement because it effectively forces one language for all readers and users of the skill guidance.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly instructs automatic collection from multiple external sources, analysis against user memory, pushing to all configured channels, and long-term storage, but it does not require clear user notice or consent for what data is collected, retained, and redistributed. This creates a privacy and autonomy risk because sensitive interests, behavioral signals, or memory-derived inferences could be processed and propagated without the user understanding the scope.

Content

No source excerpt is available for this finding.