Back to skill

Security audit

Custodian

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed OpenClaw maintenance agent, but it requests broad persistent repair authority, automatic self-updates, cross-skill scheduler changes, and automatic external searches that need careful review.

Install only if you want an autonomous OpenClaw maintenance agent that can persist across sessions, modify scheduler state, initialize other skills, repair logs/configuration, and self-update from GitHub. Before enabling it, require opt-in background tasks, disable or approve each self-update, restrict cross-skill reads and task registration, sanitize or approve web-search queries, and ensure repair commands validate placeholders and paths.

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 (3)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:64
Finding

Autonomous Cross-Skill Access and Global Scheduler Modification

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
references/known_issues.json:2
Finding

Unvalidated Placeholder Interpolation in Autonomous Repair Commands

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:177
Finding

Potential Disclosure of Sensitive Log Content Through External Web Searches

Content
View full analysis
= 1`, run next mutation query (see Web Search Protocol). ``` ### Technical Analysis The first two query mutations directly embed `{error}` into an external web-search query. Gateway and application errors can contain sensitive material, including: - Authentication tokens or authorization headers - URLs containing credentials or signed parameters - User messages or request payload fragments - Local filesystem paths and usernames - Internal hostnames, ports, and service identifiers - Email addresses or other personal information No mandatory sanitization, secret scanning, data minimization, user consent, or approved-search-provider restriction is specified before transmission. The search operation is triggered automatically after recurrence, so disclosure does not require an explicit user request for external research. Fingerprinting an error locally does not make it safe to transmit if the fingerprint or query still includes raw variable content. The protocol must explicitly distinguish a normalized error signature from the original log line. ### Attack Path 1. A gateway or Skill produces an unknown err ...[truncated 1117 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (12)

Self-Modification

High
Category
Rogue Agent
Confidence
99% confidence
Finding

A scheduled self-update mechanism from GitHub is a genuine self-modification capability because it allows the skill to replace or alter its own code on a recurring unattended basis. In the context of a tool that also registers tasks and performs automatic repairs, this meaningfully expands the blast radius of any repository compromise, dependency hijack, or logic bug into persistent autonomous execution.

Content

Scanner excerpt · README.md (reported line 48)May include surrounding context.

md
|---|---|---|---|
| `custodian:light` | heartbeat | every heartbeat cycle | `custodian.scan.light` |
| `custodian:deep` | cron | optimized 6h (initial: `0 1,7,13,19 * * *` PT) | `custodian.scan.deep` |
| `custodian:update` | cron | `0 0 * * *` (midnight daily) | Self-update from GitHub source |

## Changelog

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The description presents the skill as autonomously repairing failures during quiet hours, but does not clearly warn that it may modify cron registrations, create files/directories, rotate logs/JSONL, refresh tokens, adjust schedules, or perform external web searches. Omitting those disclosures is dangerous because users or orchestrators may grant it access under the mistaken belief that it is primarily observational.

Content

No source excerpt is available for this finding.

Self-Modification

High
Category
Rogue Agent
Confidence
97% confidence
Finding

The skill schedules a self-update from a GitHub source, which introduces a supply-chain and remote code change risk: future behavior can change without fresh user review. Combined with autonomous background execution, this can allow unreviewed updates to gain ongoing privileged influence over maintenance actions.

Content

Scanner excerpt · SKILL.md (reported line 186)May include surrounding context.

md
|---|---|---|---|
| `custodian:light` | heartbeat | every heartbeat cycle | `custodian.scan.light` |
| `custodian:deep` | cron | optimized 6h (initial: `0 1,7,13,19 * * *` PT) | `custodian.scan.deep` |
| `custodian:update` | cron | `0 0 * * *` (midnight daily) | Self-update from GitHub source |

Registration during `custodian.init` (idempotent -- check `openclaw cron list` first).

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The opening description explicitly says the skill detects and repairs failures autonomously during quiet hours, but it presents this as a feature without a strong warning about unattended remediation. Autonomous repairs can alter system configuration, logs, task registration, or skill state without supervision, increasing the risk of unintended changes or abuse if repair logic is wrong or manipulated.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The README states that initialization runs automatically and creates directories, config files, background tasks, and copies a workflow plan, but it does not clearly warn users that invoking the skill changes system state. Hidden or under-emphasized automatic system modifications reduce informed consent and can lead users to install or invoke the skill without understanding that persistent scheduled execution will be registered.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The scheduled tasks table documents a nightly self-update from GitHub, but the README does not present this as a significant security-sensitive behavior. Automatic code updates from a remote source can introduce supply-chain compromise, unexpected behavior changes, or malicious updates without user review, especially when combined with background execution and repair authority.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad and operationally generic, such as 'check system health' or 'fix log errors,' which could cause the skill to be invoked in contexts where the user did not intend autonomous repair behavior. Because this skill can register tasks, modify cron state, initialize skills, and perform repairs, accidental invocation materially raises the risk of unauthorized or surprising changes.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
92% confidence
Finding

The skill is designed for persistence: it creates storage, registers recurring background tasks, writes journals, copies plans into another component's directory, and continuously re-runs maintenance logic. Persistent autonomous execution increases blast radius because any misconfiguration, unsafe repair rule, or compromised update can continue acting without immediate user awareness.

Content

Scanner excerpt · SKILL.md (reported line 35)May include surrounding context.

md
## Commands

- `custodian.init` -- create storage, register background tasks idempotently, copy bundled plan to Mentor plans dir, build initial activity model
- `custodian.scan.light` -- tail gateway log (last 100 lines), check cron registry, retry open `fix_attempted_failed` items, check for uninitialized skills, write Observation Journal
- `custodian.scan.deep` -- full sweep: all light steps + full JSONL scan, doctor diagnostic, journal health, skill conformance, skill init pass, activity model rebuild, schedule optimization, repair pass, web search pass, escalation pass, report, Vesper signal
- `custodian.verify {fix_id}` -- verify fix outcome, update record, escalate if failed twice

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill claims a non-destructive, reversible safety envelope, but elsewhere explicitly directs state-changing actions such as rotating files, generating default configs, re-enabling cron jobs, registering tasks, and updating schedules. This mismatch can mislead users or downstream agents into authorizing autonomous modifications under a false assumption of harmlessness, increasing the chance of unintended system changes.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
80% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/known_issues.json (reported line 125)May include surrounding context.

json
"source": "doctor",
    "auto_fix": null,
    "reversibility": null,
    "description": "systemd linger not enabled -- background services will not persist across logout",
    "technology_context": "systemd loginctl linger"
  },
  "oc_node_version_mismatch": {

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill advertises very broad trigger phrases such as 'check system health', 'why is X failing', and 'clean up errors', which are likely to match common user requests outside the intended scope. Because this skill is described as autonomously applying fixes, initializing skills, and registering background tasks, accidental invocation could lead to unintended operational changes rather than merely returning information.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

This markdown file explicitly instructs the skill to write a proposal to ~/openclaw/data/ocas-custodian/reports/repair-{{params.issue_id}}.md, but it does not include any user-facing warning that the plan will create or overwrite a file on disk. Under the markdown criteria for SQP-2, behaviors affecting user data or system state should be disclosed in the skill description.

Content

No source excerpt is available for this finding.