Back to skill

Security audit

Evomasterscheduler

Security checks for vulnerabilities and agentic risk

Overview

This skill needs Review because it installs unattended maintenance jobs that copy secrets, mutate repositories, delete logs, restart services, and report operational details to Slack without enough scoping or opt-in.

Install only after reviewing every scheduled job. Treat it as granting recurring access to your OpenClaw config, workspace, agents, logs, gateway controls, Git repository, and Slack integration. Disable or edit the backup, Git commit, log deletion, and Slack notification behavior unless you explicitly want them, and do not let it copy .env files without encryption and retention controls.

Vulnerability Patterns
  • 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
  • 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 (5)

T06 · System Persistence

Error
Location
consolidate-crontab.sh:30
Finding

Unattended Cross-Session Persistence Through Scheduled Tasks

Content
View full analysis
&1 || echo "Cron 1 creation logged" openclaw cron add --name evomaster-fix --cron "1 3 * * *" --message "Run: $SCRIPT_DIR/cron-fix.sh" --description "EvoMaster: Self-Fixing" 2>&1 || echo "Cron 2 creation logged" openclaw cron add --name evomaster-memory --cron "2 3 * * *" --message "Run: $SCRIPT_DIR/cron-memory.sh" --description "EvoMaster: Memory Maintenance" 2>&1 || echo "Cron 3 creation logged" # CRON 4: 3:05 AM (backup) openclaw cron add --name evomaster-backup --cron "5 3 * * *" --message "Run: $SCRIPT_DIR/cron-backup.sh" --description "EvoMaster: Self-Backing-Up" 2>&1 || echo "Cron 4 creation logged" # CRON 5: 3:10 AM (improve) openclaw cron add --name evomaster-improve --cron "10 3 * * *" --message "Run: $SCRIPT_DIR/cron-improve.sh" --description "EvoMaster: Self-Improving" 2>&1 || echo "Cron 5 creation logged" # CRON 6: 3:15 AM (upgrade) openclaw cron add --name evomaster-upgrade --cron "15 3 * * *" --message "Run: $SCRIPT_DIR/cron-upgrade.sh" --description "EvoMaster: Self-Upgrading" 2>&1 || echo "Cron 6 creation logged" # CRON 7: 3:20 AM (security) openclaw cron add --name evomaster-security --cron "20 3 * * *" --message "Run: $SCRIPT_DIR/cron-security.sh" --description "EvoMaster: Security Audit" 2>&1 || echo "Cron 7 creation logged" ``` The installation instruction explicitly directs the user to register these jobs: ```markdown ## Installation Run `consolidate-crontab.sh` to synchronize these tasks with the system crontab. ``` ### Technical Analysis The instal ...[truncated 1992 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
cron-backup.sh:11
Finding

Plaintext Credential Duplication in Broad Unprotected Backups

Content
View full analysis
/dev/null || true cp ~/.openclaw/.env "$BACKUP_DIR/" 2>/dev/null || true # Backup workspace cp -r ~/.openclaw/workspace "$BACKUP_DIR/workspace/" 2>&1 # Backup agents config cp -r ~/.openclaw/agents "$BACKUP_DIR/agents/" 2>&1 ``` ### Technical Analysis The scheduled backup explicitly copies `~/.openclaw/.env`, which is a credential-related file, into a date-based directory under `~/Backups`. It also recursively copies the complete workspace and agents directory, which may contain additional confidential state. The script does not set a restrictive `umask`, enforce directory mode `0700`, enforce file mode `0600`, encrypt the backup, exclude secrets, or implement retention. Consequently, backup permissions depend on the user's ambient umask and existing destination state. The operation creates another long-lived plaintext copy of credentials and increases the number of locations an attacker must compromise. Credential content is not transmitted directly by this code, but local duplication is unnecessary for many configuration-backup use cases and exceeds minimum privilege unless secret backup is explicitly requested. ### Attack Path 1. The persistent backup job runs each day. 2. The script creates a date-based backup directory using ambient filesystem permissions. 3. It copies `.env`, configuration, workspace content, and agent state into that directory. 4. Another local account, compromised process, synchronization client, or backup collector with access to `~/Backups` reads the duplicated `.env`. 5. Exposed tokens or credentials are then used against the services for which they are valid. ### Impact Assessment An attacker ...[truncated 431 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
cron-backup.sh:28
Finding

Automatic Git Commit Can Persist Secrets and Unintended Workspace Data

Content
View full analysis
/dev/null || true ``` ### Technical Analysis The backup job changes into the entire OpenClaw workspace and runs `git add .`, staging every modified or untracked path not excluded by repository rules. It performs no path allowlisting, secret scan, diff review, repository-remote check, or user confirmation. Git history is persistent: removing a sensitive file in a later commit does not remove it from earlier objects. Although this script does not itself push to a remote, a later routine or manual push can publish every automatically committed file. This behavior also mutates the user's source repository even though the declared task is a filesystem backup. The trailing `|| true` suppresses failures, while later notification text unconditionally says that a Git commit was applied. This can conceal repository errors and mislead the operator about backup status. ### Attack Path 1. A credential, private document, generated artifact, or other sensitive file appears anywhere under `~/.openclaw/workspace`. 2. The daily backup job executes `git add .`. 3. The sensitive content is written into local Git history by the automatic commit. 4. The repository is later pushed, mirrored, shared, or inspected by another user or automation process. 5. The sensitive content remains recoverable from Git history even after deletion from the working tree. ### Impact Assessment The direct impact is unauthorized persistence of arbitrary workspace content in repository history. Subsequent publication can disclose credentials, proprietary data, prompts, logs, or agent state to anyone with access to the remote or copied repository. The operation has the user's repository-write p ...[truncated 73 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
cron-security.sh:21
Finding

Security Audit Reports Unsupported Permission and Exposure Assurances

Content
View full analysis
> "$LOG_FILE" else echo "[ERROR] $(basename $file) is not readable." >> "$LOG_FILE" fi done # Send Slack notification bash "$HOME/.openclaw/scripts/slack-notify.sh" "$SLACK_CHANNEL" "🔒 **EvoMaster Security Audit Complete** ($TIMESTAMP) $SECURITY_STATUS ✅ File permissions verified ✅ Gateway binding verified (localhost) ✅ No credential exposure ``` ### Technical Analysis The file check tests only whether the current execution user can read each file. Readability by the intended owner is not evidence of secure permissions. A world-readable `.env`, a file owned by an unexpected user, or a file writable by another account will still pass this test. The script does not inspect ownership, group membership, numeric permission modes, access-control lists, symlink targets, or parent-directory permissions. It also does not inspect the gateway's listening interfaces or search for credential exposure. Nevertheless, it unconditionally sends affirmative claims that file permissions, localhost binding, and credential isolation were verified. This creates a false-negative security control: insecure configurations can be presented as safe to operators. ### Attack Path 1. `.env` or `openclaw.json` receives unsafe permissions, ownership, or a malicious symlink target, or the gateway becomes externally bound. 2. The scheduled security task runs under the normal user. 3. The `-r` test succeeds because that user can read the file. 4. No gateway-binding or credential-exposure validation occurs. 5. The notification states that permissions, localhost binding, an ...[truncated 527 chars]
Remediation
View remediation

other

Warning
Location
cron-backup.sh:31
Finding

Operational Metadata Is Sent to Hardcoded Slack Channels Without Documentation or Opt-In

Content
View full analysis
/dev/null || echo "Slack notification failed" ``` From `cron-upgrade.sh:25-38`: ```sh # Send Slack notification bash "$HOME/.openclaw/scripts/slack-notify.sh" "$SLACK_CHANNEL" "⬆️ **EvoMaster Self-Upgrading Check** ($TIMESTAMP) Current Version: $current_version Status: $UPGRADE_STATUS 📋 To upgrade manually: • OpenClaw: \`openclaw update\` • Ollama: \`ollama pull \` • Nerve: Rebuild from source 📄 Full log: \`$LOG_FILE\` *Next run: Tomorrow 3:15 AM*" ``` ### Technical Analysis Nearly every maintenance script sends information through the configured Slack integration to a hardcoded channel name. The Skill documentation describes local logging but does not disclose external reporting, enumerate transmitted fields, or provide an opt-in setting. Reported information includes local backup and log paths, backup size, disk usage, gateway status, software version, error and warning counts, remediation status, security status, memory synchronization state ...[truncated 1267 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (15)

Credential Access

High
Category
Privilege Escalation
Confidence
98% confidence
Finding

Copying ~/.openclaw/.env into backups is a direct credential-handling risk because .env files commonly store API keys, tokens, passwords, and other secrets. Duplicating secrets into another location expands the attack surface, increases retention of sensitive data, and makes compromise more likely if backups are less protected than the source.

Content

Scanner excerpt · cron-backup.sh (reported line 16)May include surrounding context.

sh
# Backup core config files
cp ~/.openclaw/openclaw.json "$BACKUP_DIR/" 2>/dev/null || true
cp ~/.openclaw/.env "$BACKUP_DIR/" 2>/dev/null || true

# Backup workspace
cp -r ~/.openclaw/workspace "$BACKUP_DIR/workspace/" 2>&1

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · cron-security.sh (reported line 24)May include surrounding context.

sh
# Verify file permissions
CRITICAL_FILES=(
    ~/.openclaw/openclaw.json
    ~/.openclaw/.env
)

for file in "${CRITICAL_FILES[@]}"; do

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill describes multiple automated tasks that can modify system state, including remediation, backups, upgrades, security changes, and crontab synchronization, but does not warn the user about those effects or require explicit confirmation. In an agent-skill context, this lack of disclosure can lead operators to install or enable the skill without understanding that it may persist scheduled jobs and make autonomous changes to the host environment.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The script announces 'Consolidate: Remove duplicate/legacy crons' and later says an old cron is being 'marked only' for removal, which implies cleanup-focused behavior. In practice, the code from L30-L50 only adds new cron entries and never removes or marks existing ones programmatically, so the documentation materially misstates the script's operational effect.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The header frames the script as a simple self-backup job, but the code also performs additional side effects: committing workspace changes to Git and sending an external Slack notification. This mismatch can mislead operators about the script’s real behavior, reducing informed consent and increasing the chance that sensitive changes or metadata are propagated unexpectedly.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
76% confidence
Finding

The script creates a dated backup directory and stores agent state, configuration, and workspace contents there, which establishes persistent copies of operational data. While persistence is expected for backups, in a security review this still increases exposure because compromised or overly permissive backup storage can preserve sensitive material long after it is removed from the primary location.

Content

Scanner excerpt · cron-backup.sh (reported line 12)May include surrounding context.

sh
echo "[$TIMESTAMP] [BACKUP] Starting state backup..." >> "$LOG_FILE"

BACKUP_DIR="$HOME/Backups/openclaw/$(date +%Y-%m-%d)"
mkdir -p "$BACKUP_DIR"

# Backup core config files
cp ~/.openclaw/openclaw.json "$BACKUP_DIR/" 2>/dev/null || true

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The script copies configuration, workspace, agents, and especially the .env file into persistent backup storage without any access control, encryption, or user-facing warning. These files can contain credentials, tokens, prompts, proprietary data, and operational state, so a compromise of the backup location exposes a broad set of sensitive material.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The script sends backup metadata to Slack, an external service, without any explicit warning or consent flow in the script. Even if only metadata is transmitted, backup paths, timestamps, operational cadence, and log locations can reveal environment structure and create useful reconnaissance information for an attacker or overexpose internal operations to third parties.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

This code automatically detects a down process and issues openclaw gateway start, which changes system state by restarting a service. Although the action is logged, there is no user-facing warning or confirmation, and the script itself does not explain this safety-impacting behavior beyond an internal comment.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The Slack message reports multiple security checks as passed, including file permissions, gateway binding, and credential exposure, even though the script only checks whether two files are readable and does not validate the other claims at all. This creates a false sense of security and can cause operators to miss real misconfigurations or incidents because the notification is presented as an authoritative audit summary.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The script automatically transmits security-audit status and the full local log path to Slack, which exposes operational security information to a third-party messaging system and to anyone with channel access. While the message shown here does not include raw secrets, security findings and environment details may still be sensitive and could aid reconnaissance or broaden exposure of internal audit results.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The top-level comment labels the script as 'Self-Upgrading', suggesting it automatically upgrades components. In reality, the code logs version information, lists Ollama models, and explicitly tells the user that manual upgrade is required, so the documentation contradicts the actual behavior.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

With no manifest available, the only stated purpose comes from the file comments and inline labels describing 'Self-Fixing' and automated remediation. The script not only checks and restarts OpenClaw, but also invokes a separate Slack notification script to send status externally, which is a distinct communications capability not justified by the limited repair behavior documented in this file.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
72% confidence
Finding

The script sends a Slack message containing the full log file path, which is system metadata transmitted to an external communication channel. While the behavior is visible in code, there is no user-facing disclosure in the script comments or documentation warning that operational details are sent to Slack.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

This code transmits system status information, including the current OpenClaw version and local log file path, to an external Slack notification script. While the behavior is commented as 'Send Slack notification,' there is no user-facing disclosure, confirmation, or warning in the script itself about sending this data off-host.

Content

No source excerpt is available for this finding.