Back to skill

Security audit

Model Registry Manager

Security checks for vulnerabilities and agentic risk

Overview

The skill does model-registry maintenance, but it also creates persistent jobs, uses API keys against unvalidated endpoints, and can overwrite live OpenClaw configuration without enough guardrails.

Review before installing. Use this only in a controlled OpenClaw environment with backed-up configuration, trusted provider base URLs, scoped and revocable API keys, and explicit approval for any cron jobs, `.learnings/` writes, config changes, or gateway restarts.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (5)

T06 · System Persistence

Error
Location
SKILL.md:92
Finding

Automatic Creation of Persistent Scheduled Jobs with Command Execution

Content
View full analysis
Remediation
View remediation

T02 · Agent Memory Poisoning

Warning
Location
SKILL.md:64
Finding

Unprompted Writes to Persistent Agent Learning State

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/sync-models.js:44
Finding

Synchronization Script Sends API Credentials to Unvalidated Provider Endpoints

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/validate-model-sync.js:26
Finding

Validation Script Sends API Credentials to Unvalidated Provider Endpoints

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/sync-models.js:118
Finding

Partial or Failed Synchronization Can Destructively Erase Model Catalog Entries

Content
View full analysis
[m.id, m])); const remote = await fetchModels(provider); const dedup = new Map(); const duplicateRemoteIds = []; const denied = []; for (const m of remote) { if (!m?.id) continue; if (!isAllowed(policy, m.id)) { denied.push(m.id); continue; } if (dedup.has(m.id)) { duplicateRemoteIds.push(m.id); continue; } dedup.set(m.id, m); } const verified = []; const unusable = []; const ordered = [...dedup.values()]; const limit = VERIFY_LIMIT > 0 ? Math.min(VERIFY_LIMIT, ordered.length) : ordered.length; for (const model of ordered.slice(0, limit)) { const ok = await probeModel(provider, model.id).catch(() => false); if (!ok) { unusable.push(model.id); continue; } const normalized = normalizeModel(model, previousModels.get(model.id)); verified.push(normalized); allVerified.push({ providerKey, ...normalized }); } provider.models = verified; providerReports.push({ provider: providerKey, fetched: rem ...[truncated 2209 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (14)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The documented purpose says the skill performs safe registry synchronization, but the behavior described includes active probing of provider endpoints with API-backed access and scheduled operational validation. This mismatch can mislead operators into approving or invoking a skill under narrower assumptions than its real behavior, causing unreviewed external calls and persistent automation.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 95)May include surrounding context.

md
- `SKILL.md`: only workflow + decision rules

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
88% confidence
Finding

The skill instructs network access and operational actions against provider endpoints, but it does not declare an explicit tool scope such as allowed tools or permissions. That creates an authorization gap where an agent may run with broader-than-necessary capabilities, increasing the chance of unintended network access or access to sensitive environment-backed credentials.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The invocation guidance is broad enough to trigger the skill in general model-management situations, even when persistent changes, probing, or scheduling may be inappropriate. Over-broad activation criteria increase the chance of the agent performing network calls, config edits, or cron setup in contexts where the user only wanted inspection or advice.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill directs automatic persistence of run-derived information into .learnings/ as a default closing step, without requiring explicit user consent or a sensitivity check. This can store user corrections, operational failures, or workspace-specific details persistently, creating data leakage, retention, and prompt-injection persistence risks across future runs.

Content

No source excerpt is available for this finding.

Persistent Context Injection

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill injects content designed to persist in agent memory or context across interactions. Persistent injection can alter agent behavior long after the initial interaction.

Content

Scanner excerpt · SKILL.md (reported line 89)May include surrounding context.

md
- per-provider sync + global catalog merge

### Avoid hardcoding
- exact vendor model ids as permanent rules
- one provider’s naming quirks as universal truth
- one-off emergency fixes in the main sync algorithm
- model routing policy into registry sync

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill recommends creating or updating recurring cron jobs, which is a persistent system or workspace change, but it does not require an explicit warning and confirmation before doing so. Persistence plus scheduled execution can repeatedly trigger network access, script execution, and configuration changes long after the original interaction, making mistakes or abuse much more damaging.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The scheduled job guidance allows broad command execution via toolsAllow including exec for a task whose purpose is limited to model sync and validation. Granting exec to recurring jobs expands the blast radius of any prompt mistake, script compromise, or downstream misuse into arbitrary command execution on every scheduled run.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The script overwrites the primary OpenClaw configuration file directly when not run with --dry-run, with no confirmation prompt, backup, or safer staged write. Because the file controls provider endpoints, API-backed model registration, and downstream agent defaults, an operator can unintentionally destroy or replace working configuration state, especially if remote provider data is incomplete or maliciously manipulated.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

When --restart is supplied, the script immediately runs 'openclaw gateway restart' after changing configuration, without a confirmation step or guardrail. In this skill's context, the script is modifying live model registry state based on remote provider responses, so an unsafe or accidental run can trigger service disruption or apply bad config directly to production-facing components.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This code performs HTTP requests to external provider endpoints using an Authorization bearer token and sends model IDs in request bodies, but the file contains no docstring, comment, or user-facing disclosure explaining that network calls and credential-backed probes will occur. For a code file, these are safety-relevant operations involving external transmission of system and configuration data.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

The skill explicitly states it should stay focused on model registry sync and should not bundle broader operational policy, yet later sections prescribe workspace scheduling, job structure, delivery mode, and failure alert behavior. These instructions expand the skill into ongoing operational orchestration, which conflicts with the earlier documented boundary that the skill remain focused on registry sync and validation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
97% confidence
Finding

The markdown specifies 'Recommended default schedules (Asia/Shanghai)', which hard-codes a locale/timezone preference. This is a natural-language locale policy choice presented as the default without offering the user a choice or explaining why that region is required.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
85% confidence
Finding

The script creates a reports directory and writes JSON and Markdown report files, but there is no comment, docstring, or other in-file disclosure that local filesystem state will be modified. This meets the code-file criterion for file writes lacking any form of user warning or explanation.

Content

No source excerpt is available for this finding.