Back to skill

Security audit

优先级教练 · Priority Coach

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because it reports capabilities beyond personal coaching, including project snapshots, release analytics, and an unsafe record deletion path.

Consider avoiding installation until the extra maintenance tools are clearly separated or their scope and permissions are reviewed, and the record-deletion confirmation issue is addressed. Skillspector inspected all listed components but marked its analysis partial because two references could not be resolved, so the evidence is incomplete.

Vulnerability Patterns
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/record.py
Finding

Record deletion bypasses the documented consent gate

Content
View full analysis

Impact

A direct helper invocation can permanently remove saved coaching history or session summaries without the promised confirmation gate. This is an unsafe implementation, not evidence of deliberate malicious behavior.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (24)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied code has a materially different primary purpose from the declared coaching skill: it monitors software iterations and adoption metrics for a maintainer. Its project-wide snapshots, Git inspection, optional URL fetching, and analytics reports are not implementation details needed to select personal priorities. These capabilities are absent from the description. The script acts through explicit CLI subcommands; the code does not demonstrate automatic execution or transmission of captured project contents.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied script implements release analytics rather than personal-priority coaching. Its file snapshots, Git inspection, download fetching, and growth reporting are not described and are not necessary implementation details of the declared coaching function. Although located under scripts/dev and apparently intended as a maintainer utility, its primary purpose materially differs from the declared purpose. The code does not show automatic execution during coaching or transmission of snapshot contents to a remote service; the mismatch is the separate, undeclared monitoring capability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The report reviews the coaching software's performance rather than the user's weekly priorities. Its main outputs identify product failures, recommend implementation changes, and suggest software releases—materially different from the declared personal-coaching purpose. Reading local session history could support continuity or a personal weekly review, but the maintenance-focused analysis goes beyond those supporting details. The code does not transmit data, invoke an AI service, modify the skill's source, or calculate working hours; it only generates reports and prompts, optionally writing them to supplied local paths.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The supplied command concerns release-download data rather than helping a user prioritize personal activities. This is an undeclared operational capability with no evident supporting role in the stated coaching function. The snippet does not establish the precise data fetched, network destinations, or storage behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied command initiates release-monitor reporting rather than prioritization coaching. No connection to the declared coaching workflow is shown. The script implementation is not supplied, so its precise reporting behavior and any resource access cannot be established.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied command performs an operational release-monitoring action rather than providing prioritization guidance. Nothing in the declared purpose explains this functionality or establishes it as supporting implementation. The script itself is not supplied, so its exact data access, network activity, and side effects cannot be determined.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied command concerns software release tracking rather than helping a person select priorities or begin a small action. It appears to be a maintenance operation, but the visible chunk does not connect it to implementing the declared coaching behavior. The script body is not supplied, so no additional file access, network activity, or data collection can be established.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The declared weekly reflection concerns the user's priorities, whereas this script reviews the coaching product itself and recommends maintenance work and releases. That is a materially different purpose, beyond merely storing or retrieving sessions to support coaching. Reading local session data is consistent with continuity and is not independently suspicious. The script performs no network transmission, invokes no AI service, and does not itself modify the skill's implementation; it prints reports or writes them to explicitly supplied output paths.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The visible commands expose a release-reporting workflow rather than coaching someone on what to prioritize next. Although the description mentions reviewing the week's main focus, nothing connects that coaching function to release monitoring. The mismatch is based on the visible command interface; network access, data sources, and other internal behavior cannot be established from this snippet.

Content

No source excerpt is available for this finding.

Code capability 'network' detected in scripts/dev/release_monitor.py but not covered by declared permissions.

High
Category
MCP Least Privilege
Confidence
83% confidence
Finding

The optional fetching path passes a CLI- or SKILL.md-supplied URL to urlopen without restricting schemes, destinations, or redirects. If an agent runs it using attacker-controlled metadata, it can access internal services or local resources with the agent's privileges; explicit invocation and the development-only context reduce exposure, and no automatic upload of coaching records is shown.

Content

No source excerpt is available for this finding.

Code capability 'env' detected in scripts/dev/release_monitor.py but not covered by declared permissions.

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'env' capability that is not covered by the declared 'allowed-tools' preapproval. Review the host runtime's actual enforcement policy; runtime permissions are unknown unless separately verified, and this declaration alone does not establish a permission ceiling.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx clawhub'.

Medium
Category
MCP Rug Pull
Confidence
97% confidence
Finding

The publishing command uses npx clawhub@latest, which executes a mutable npm package version rather than a reviewed, fixed dependency. A compromised future release could execute code with the maintainer's filesystem access and publishing credentials; this is a maintenance supply-chain risk, not an unpinned MCP server.

Content

No source excerpt is available for this finding.

Generic return phrases can activate the priority-review workflow without establishing the user's intended context.

Medium
Category
Trigger Abuse
Confidence
93% confidence
Finding

The activation examples include “我回来了” (“I'm back”) and “接着上次” (“continue from last time”), which commonly occur outside priority planning. Requiring user initiation does not disambiguate these phrases, so an ordinary greeting or request to resume another topic could trigger this workflow.

Content

No source excerpt is available for this finding.

The local-record workflow instructs immediate script execution without warning about data-changing effects.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

This workflow explicitly directs the agent to invoke record.py for intents that include saving and deleting records, rather than merely listing available tools. It provides no user-facing disclosure of what will be stored or which records deletion will affect before execution.

Content

No source excerpt is available for this finding.

Everyday statements are treated as state triggers without sufficient contextual constraints.

Medium
Category
Trigger Abuse
Confidence
90% confidence
Finding

Triggers such as “我回来了” (“I'm back”) and “我太累了” (“I'm very tired”) can occur without a request to resume coaching or enter an overwhelmed-state workflow. Although this is a coaching router, these triggers do not require relevant prior coaching context or distinguish casual statements from requests to change state.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/dev/release_monitor.py (reported line 112)May include surrounding context.

python
def run_git(project_root: Path, args: list[str]) -> bytes:
    try:
        completed = subprocess.run(
            ["git", *args],
            cwd=project_root,
            check=True,

Project-wide source snapshotting is unrelated to the declared personal-coaching purpose.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

Worktree snapshotting reads every small UTF-8 file beneath the selected root, without excluding secrets or checking file symlinks, and persists complete contents in JSON snapshots. This can duplicate .env files, private data, or symlinked external files into retained monitoring artifacts; it is a local disclosure and retention risk rather than demonstrated network exfiltration.

Content

No source excerpt is available for this finding.

Download-count scraping introduces network access unrelated to personal priority coaching.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

Being unrelated to coaching is not itself a vulnerability, and download monitoring is a plausible maintainer feature. However, this path actually fetches an unchecked homepage from skill metadata, creating a conditional attacker-directed request risk when invoked by an agent with access to internal resources; it is the same underlying network issue identified by LP1.

Content

No source excerpt is available for this finding.

Tainted flow: 'legacy' from os.environ.get (line 711, credential/environment) → shutil.copy2 (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/record.py (reported line 718)May include surrounding context.

python
if not legacy.exists():
            write_store(kind, empty_store(kind))
            return f"{kind}: 未发现旧数据,已初始化新路径:{primary}"
        shutil.copy2(legacy, primary)
        data, _ = load_store(kind)
        if items_key not in data:
            data[items_key] = []

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/release_monitor.py (reported line 112)May include surrounding context.

python
def run_git(project_root: Path, args: list[str]) -> bytes:
    try:
        completed = subprocess.run(
            ["git", *args],
            cwd=project_root,
            check=True,

Project-wide file collection and persistent source snapshots are unrelated to personal priority coaching.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Worktree capture collects nearly every small UTF-8 file under the project, including untracked .env files, credentials, or personal notes, and persists complete contents in snapshots. It also follows file symlinks without checking that their targets remain inside the project, potentially copying outside files into the monitoring store. Release monitoring explains the functionality, but the broad collection and lasting copies create unnecessary confidentiality risks for a personal coaching skill; no network upload is shown.

Content

No source excerpt is available for this finding.

Fetching external download metrics introduces an undeclared release-analytics network capability.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

When explicitly invoked with --fetch, the script passes a CLI URL or skill-controlled homepage to urlopen without restricting schemes, destinations, or redirects. An attacker who controls that metadata can cause requests to internal services or local-file URLs under the runner's privileges, although the code stores only a parsed download count rather than the complete response. Fetching public metrics is a plausible maintenance feature, but the missing destination checks create a conditional request-forgery risk unrelated to coaching.

Content

No source excerpt is available for this finding.

The morning_start activation signals include everyday phrases without an explicit planning-context constraint.

Low
Category
Trigger Abuse
Confidence
91% confidence
Finding

The listed signals include “我准备开始了” (“I'm ready to start”) and “我有点拖” (“I'm procrastinating a bit”), which can occur in many unrelated conversations. The activation guidance does not require an established daily plan or exclude unrelated uses, so these phrases could unintentionally select the morning-start workflow.

Content

No source excerpt is available for this finding.

The report generator implements skill-maintenance and release-planning capabilities outside the personal coaching scope.

Low
Category
Excessive Agency
Confidence
93% confidence
Finding

The manifest describes helping an individual identify priorities and review their weekly focus. Instead, this prompt requests diagnosis of internal router/state components, release decisions, files to modify, and a changelog; the report also computes version-bump recommendations at L239–L244. These are product-maintenance capabilities rather than personal coaching, although this script only generates recommendations and does not execute changes.

Content

No source excerpt is available for this finding.