Back to skill

Security audit

Watchitai

Security checks for vulnerabilities and agentic risk

Overview

WatchItAI clearly describes screen sharing and remote control, but it runs unverified downloaded programs with screen and control permissions, so it needs careful review before installation.

Install only if you trust WatchItAI and are comfortable granting screen-sharing and optional remote-control access. Prefer a release with signed or hashed binaries, use view-only mode unless control is necessary, avoid sharing sensitive screens, and remove the skill directory plus config.json/accessKey when you no longer need it.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (4)

T03 · Remote Payload Retrieval and Execution

Error
Location
run.sh:50
Finding

Unverified Remote Executable Download and Immediate Execution

Content
View full analysis
&2 echo " $url" >&2 mkdir -p "$BIN_DIR" if command -v curl >/dev/null 2>&1; then curl -fSL --connect-timeout 15 -o "$target.tmp" "$url" 2>/dev/null && mv "$target.tmp" "$target" && return 0 elif command -v wget >/dev/null 2>&1; then wget -q --timeout=15 -O "$target.tmp" "$url" 2>/dev/null && mv "$target.tmp" "$target" && return 0 fi rm -f "$target.tmp" 2>/dev/null echo "❌ Download failed: $url" >&2 return 1 } if [ ! -f "$BINARY_PATH" ]; then if [ -f "$GZ_PATH" ]; then decompress_binary "$GZ_PATH" "$BINARY_PATH" || exit 1 else echo "🔧 Binary not found locally, attempting download..." >&2 if download_binary "$BINARY_NAME"; then chmod +x "$BINARY_PATH" 2>/dev/null || true macos_sign_binary "$BINARY_PATH" else # Platform fallback omitted here only after the same download # function is invoked for the fallback binary. exit 1 fi fi fi exec "$BINARY_PATH" "$@" ``` The Windows launcher implements the same trust model: ```cmd set "DOWNLOAD_URL=https://watchitai.net/downloads/%BINARY_NAME%" powershell -NoProfile -Command "try { [Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; Invoke-WebRequest -Uri '%DOWNLOAD_URL%' -OutFile '%BINARY_PATH%' -UseBasicParsing -TimeoutSec 60 } catch { Write-Host '❌ Download failed: ' $_.Exception.Message; exit 1 }" cd /d "%SCRIPT_DIR%" ...[truncated 2418 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
run.sh:53
Finding

macOS Quarantine Removal and Ad-Hoc Signing of Downloaded Executables

Content
View full analysis
/dev/null 2>&1; then xattr -d com.apple.quarantine "$target" 2>/dev/null || true codesign --force --sign - "$target" 2>/dev/null || true fi } ``` ### Technical Analysis The launcher removes the `com.apple.quarantine` extended attribute and then applies an ad-hoc signature to the target executable. This does not establish publisher authenticity. Instead, it can suppress or weaken macOS provenance checks and user warnings that would otherwise apply to a downloaded executable. The function is called for locally decompressed and remotely downloaded binaries. It does not first verify a trusted Developer ID, notarization ticket, Team ID, or cryptographic artifact digest. Consequently, an attacker-controlled payload obtained through the download channel can be prepared for execution using the same logic. Removing quarantine is not required for the declared screen-sharing functionality. Properly signed and notarized distribution would preserve the operating system’s trust boundary. ### Attack Path 1. An attacker causes a malicious binary to be returned from the configured download endpoint or replaces a packaged compressed binary before installation. 2. The launcher writes or decompresses the binary into the Skill’s `bin` directory. 3. `macos_sign_binary` deletes its quarantine metadata. 4. The launcher applies an ad-hoc signature that carries no trusted publisher identity. 5. The launcher executes the binary. 6. The malicious binary operates with the user’s privileges and any sensitive permissions granted to the process. ### Impact Assessment ...[truncated 496 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
SKILL.md:147
Finding

Mutable and Unverified Skill Archive Installation Sources

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
run.sh:134
Finding

Account Access Key Stored in a General-Purpose Configuration File Without Consistent Permission Hardening

Content
View full analysis
"${SCRIPT_DIR}/config.json" <<'EOF' { "domain": "watchitai.net", "bridgePort": 8765, "mode": "server" } EOF chmod 600 "${SCRIPT_DIR}/config.json" fi ``` The later execution context confirms that the same file is expected to hold the credential: ```bash # Run from SCRIPT_DIR so config.json (including accessKey) lives next to # the skill install dir regardless of the user's shell working directory. cd "$SCRIPT_DIR" exec "$BINARY_PATH" "$@" ``` The Windows launcher creates the file without applying a restrictive ACL: ```cmd if not exist "%SCRIPT_DIR%config.json" ( ( echo { echo "domain": "watchitai.net", echo "bridgePort": 8765, echo "mode": "server" echo } ) > "%SCRIPT_DIR%config.json" ) ``` ### Technical Analysis The launchers intentionally place an account `accessKey` in `config.json` beside the Skill files. On Unix, owner-only permissions are applied only when the launcher creates a missing file. The packaged `config.json` already exists, so this branch normally does not run during initial use, and the launcher does not enforce secure permissions on an existing file. On Windows, no explicit ACL restriction is applied. The unavailable Go binary performs the actual credential write, so atomicity, file mode preservation, logging behavior, token scope, and encryption cannot be verified. A general-purpose configuration file is also more likely to be copied into backups, synchronized, shared for troubleshooting, or accidentally committed. ### Attack Path 1. ...[truncated 1090 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 (16)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The script will fetch a platform-specific executable from a remote server when the local binary is missing, then mark it executable and immediately run it. This is dangerous because it creates a download-and-execute path outside the packaged skill contents, shifting trust to network delivery and the remote endpoint without any visible checksum, signature verification, or explicit user consent at execution time.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The README promotes real-time screen sharing and optional remote mouse/keyboard control, but it does not prominently warn users that this grants access to highly sensitive on-screen data and can allow direct system interaction. In a skill context, this is dangerous because users may install and run a prebuilt binary with elevated OS permissions for screen recording, accessibility, and input monitoring without understanding the privacy and compromise implications.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
89% confidence
Finding

The installation instructions place the skill into persistent agent skill directories and fetch a remote zip containing prebuilt binaries, making the capability available across future agent sessions. In context, persistence is more dangerous because the skill enables screen capture, screenshots, localhost bridge access, and optional remote control, so a one-time install can create ongoing exposure if later invoked by an agent or user without re-evaluating trust.

Content

Scanner excerpt · README.md (reported line 24)May include surrounding context.

Install for Trae

bash
mkdir -p ~/.trae-cn/skills && cd ~/.trae-cn/skills
curl -L https://watchitai.net/watchitai-skill.zip -o watchitai.zip
unzip watchitai.zip && rm watchitai.zip

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
85% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · README.md (reported line 40)May include surrounding context.

Install for Claude Code

bash
mkdir -p ~/.claude/skills && cd ~/.claude/skills
curl -L https://watchitai.net/watchitai-skill.zip -o watchitai.zip
unzip watchitai.zip && rm watchitai.zip

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
85% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 163)May include surrounding context.

Install for Claude Code

bash
mkdir -p ~/.claude/skills && cd ~/.claude/skills
curl -L https://watchitai.net/watchitai-skill.zip -o watchitai.zip
unzip watchitai.zip && rm watchitai.zip

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · README.md (reported line 120)May include surrounding context.

bash
# Debian/Ubuntu
sudo apt install xdotool scrot

Usage

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 267)May include surrounding context.

bash
# Debian/Ubuntu
sudo apt install xdotool scrot

Usage

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The natural-language instructions, warnings, usage guidance, and operational details are presented in Chinese throughout the markdown. This can violate a language/locale policy when the skill does not offer user opt-in, alternate language support, or a justified region-specific constraint.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
76% confidence
Finding

The skill is installed persistently into agent skill directories and is designed to download and retain executable binaries on first run, creating durable remote screen-sharing and input-control capability in the user context. In context, persistence is more sensitive than normal because the skill requests shell, network, screen capture, environment-variable access, and input control, so a later invocation could re-enable surveillance or remote control without requiring reinstallation.

Content

Scanner excerpt · SKILL.md (reported line 155)May include surrounding context.

OpenClaw

bash
mkdir -p ~/.openclaw/skills && cd ~/.openclaw/skills
curl -L https://watchitai.net/watchitai-skill.zip -o watchitai.zip
unzip watchitai.zip && rm watchitai.zip

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The script will silently fetch and execute a native executable from the network if the local binary is absent, with only a status message and no integrity verification such as a signature or pinned checksum. In the context of a remote-viewing/screen-sharing skill, this is especially dangerous because a compromised download source or network path could deliver attacker-controlled code that gains the same access as the user and potentially captures screen or session data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The script creates a persistent config.json next to the skill and explicitly notes it may contain an accessKey, but it does not provide a clear warning, consent flow, or storage guidance to the user. Persistent local credential storage increases the risk of token exposure through backups, later file access, or accidental sharing of the skill directory.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
80% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · run.sh (reported line 145)May include surrounding context.

sh
"mode": "server"
}
EOF
    chmod 600 "${SCRIPT_DIR}/config.json"
fi

# --- Special handling for authorize command with friendly help ---------------

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
93% confidence
Finding

The documentation says this is a 'Cross-platform entry point (no Node.js needed)' and describes generic architecture auto-detection, but the file is a .cmd script that only runs on Windows and invokes PowerShell for decompression and download. This is an active contradiction in the inline documentation, not just an omitted detail.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
96% confidence
Finding

The header comment states that the script auto-detects architecture and launches the correct binary, but later code sets ARCH=amd64 unconditionally. The implementation does not perform architecture detection, so the documentation overstates what the script actually does.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

This code writes a new config.json file into the skill directory if one does not already exist. Although the action is commented for developers, there is no user-facing prompt or disclosure at runtime that the script will create or modify a file on disk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
96% confidence
Finding

The authorize help path instructs users to navigate via Chinese-language labels such as '技能授权' and '生成授权码'. This assumes a specific locale/UI language and does not provide an English alternative or indicate that the instructions are locale-specific.

Content

No source excerpt is available for this finding.