Back to skill

Security audit

xclawskill

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent XClaw network client, but it asks agents to run mutable remote installers and includes credential, fund-transfer, background-listening, and self-upgrade behavior that needs careful review.

Install only after reviewing the installer source yourself. Avoid the curl | bash path, prefer a pinned release with independent signature or checksum verification, keep state files private, do not rely on XCLAW_STATE_PASSPHRASE as real encryption, use HTTPS-only trusted base URLs, and treat withdraw, marketplace, daemon/listen, and self-upgrade as high-impact actions requiring explicit human control.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
Findings (5)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:24
Finding

Mutable remote installer is executed directly through curl-to-shell instructions

Content
View full analysis
Remediation
View remediation
' tar -xzf xclawskill.tar.gz bash xclawskill-v1.5.2/install.sh ``` The actual release URL and pinned public key must be controlled and documented by the publisher. ]]>

T08 · Insecure Dependencies

Error
Location
install.sh:16
Finding

Installers clone and install unpinned remote code and dependencies without trusted integrity verification

Content
View full analysis
/dev/null; then deps_ok=1 break fi done ``` The equivalent Windows installer performs the same unpinned clone and dependency installation: ```powershell $RepoUrl = "https://github.com/qomob/xclawskill.git" $src = Join-Path $env:TEMP "xclawskill-src" if (-not (Test-Path (Join-Path $src "scripts\xclaw_skill.py"))) { Write-Host "Cloning source..." git clone --depth 1 $RepoUrl $src } python -m pip install -q -r (Join-Path $target "requirements.txt") ``` ### Technical Analysis The installers clone the repository's current default branch without selecting a fixed commit or signed tag. They then copy and execute material from that checkout without checking `SHA256SUMS`. Consequently, the code installed can differ from the version that was statically audited. The Py ...[truncated 2348 chars]
Remediation
View remediation

T03 · Remote Payload Retrieval and Execution

Error
Location
scripts/xclaw_skill.py:1067
Finding

Self-upgrade trusts checksums supplied by the same unauthenticated remote source

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/xclaw_skill.py:54
Finding

State-file passphrase protection is ineffective because the decryption key is stored with the ciphertext

Content
View full analysis
str: from cryptography.fernet import Fernet key = Fernet.generate_key() token = Fernet(key).encrypt(plain.encode("utf-8")).decode("utf-8") # The key and ciphertext are saved together; the passphrase is only an obfuscation layer. return json.dumps({"v": 1, "key": key.decode("utf-8"), "data": token}) def _decrypt_secret(payload: str) -> str: from cryptography.fernet import Fernet obj = json.loads(payload) return Fernet(obj["key"].encode("utf-8")).decrypt( obj["data"].encode("utf-8") ).decode("utf-8") ``` ```python pk_bytes = state.get("private_key_bytes") if pk_bytes and pk_bytes.startswith("{") and os.environ.get("XCLAW_STATE_PASSPHRASE"): try: pk_bytes = _decrypt_secret(pk_bytes) except Exception: pk_bytes = None ``` ```python payload_pk = pk_bytes if pk_bytes and os.environ.get("XCLAW_STATE_PASSPHRASE"): payload_pk = _encrypt_secret(pk_bytes) save_state(self.state_file, { "agent_id": self.agent_id, "api_key": self.api_key, "jwt": self.jwt, "public_key_pem": self.public_key_pem, "private_key_bytes": payload_pk, }) ``` ### Technical Analysis The value of `XCLAW_STATE_PASSPHRASE` is never used as cryptographic key material. The code only tests whether the variable is present. `_encrypt_secret` then generates a random Fernet key and stores that key in the same JSON value as the ciphertext. Anyone who obtains the state file can read the embedded `key` field and decrypt the private key without knowing the passphrase. The design therefore provides encoding or obfuscation, not meaningful confidentiality. The state file is set to mode `0600`, which is a useful access-control measure on platforms that enforce POSIX ...[truncated 1923 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/xclaw_skill.py:110
Finding

API credentials can be transmitted to arbitrary hosts or over plaintext HTTP

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Rogue AgentSelf-Modification, Session Persistence
Findings (34)

Known Vulnerable Dependency: cryptography==46.0.3 — 13 advisory(ies): GHSA-537c-gmf6-5ccf (Vulnerable OpenSSL included in cryptography wheels); CVE-2026-69247 (cryptography: PKCS#7 EnvelopedData decryption exposes a Bleichenbacher oracle th); CVE-2026-69249 (python-cryptography: Duplicate self-signed intermediates can cause exponential p) +10 more

Critical
Category
Supply Chain
Confidence
96% confidence
Finding

The pinned dependency cryptography==46.0.3 is flagged with multiple published advisories, including issues involving vulnerable bundled OpenSSL components and cryptographic oracle/validation weaknesses. In a skill that interacts with an agent network, messaging, funds withdrawal, task submission, and identity/reputation operations, weaknesses in the cryptographic layer can undermine confidentiality, integrity, and authentication across the platform.

Content

No source excerpt is available for this finding.

Chaining Abuse

High
Category
Tool Misuse
Confidence
97% confidence
Finding

curl ... | bash is a classic arbitrary-code-execution pattern: it streams unaudited remote content directly into a shell with the user's privileges, eliminating any meaningful chance to inspect what will run. In this skill's context, the danger is amplified because the README explicitly targets AI agents, which may follow the command automatically and execute attacker-controlled code on the host if the source is compromised.

Content

Scanner excerpt · README.md (reported line 20)May include surrounding context.

便捷方式(跳过校验直接执行,存在供应链风险,请自行权衡):

bash
curl -fsSL https://raw.githubusercontent.com/qomob/xclawskill/main/install.sh | bash

凭据安全提醒:

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The documented actions include self-upgrade via remote git tags, daemon/listen long-running connectivity, and local credential/state persistence utilities beyond the narrowly advertised user-facing network operations. In skill ecosystems, this kind of scope creep is dangerous because operators may grant trust for messaging or discovery while unintentionally enabling code mutation and persistent background activity.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The documented actions include self-upgrade via remote git tags, daemon/listen long-running connectivity, and local credential/state persistence utilities beyond the narrowly advertised user-facing network operations. In skill ecosystems, this kind of scope creep is dangerous because operators may grant trust for messaging or discovery while unintentionally enabling code mutation and persistent background activity.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The documented actions include self-upgrade via remote git tags, daemon/listen long-running connectivity, and local credential/state persistence utilities beyond the narrowly advertised user-facing network operations. In skill ecosystems, this kind of scope creep is dangerous because operators may grant trust for messaging or discovery while unintentionally enabling code mutation and persistent background activity.

Content

No source excerpt is available for this finding.

External Script Fetching

High
Category
Supply Chain
Confidence
98% confidence
Finding

The documentation endorses a curl-piped-to-bash installation flow, which executes remote code immediately without independent verification at the moment of install. If the remote source, transport, or upstream account is compromised, this gives an attacker arbitrary code execution on the host.

Content

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

md
| 资金转出 | `withdraw` | 发起链上提现,转出金额与目标地址不可逆 | 必须显式 `--confirm`;执行前回显地址与金额供核对;广播由平台执行器处理 |

**文件访问**:仅读写用户显式指定的状态文件、临时目录与技能安装目录,不读取其他用户文件。
**执行环境**:`verify` 等动作调用系统 Python;`curl | bash` 安装方式请先校验(见 README 安全章节)。

## Installation (agents: do this first if the skill is missing)

Chaining Abuse

High
Category
Tool Misuse
Confidence
97% confidence
Finding

The command chain combines remote content retrieval and immediate shell execution in a single pipeline, eliminating opportunities for inspection or validation. This chaining materially increases exploitability compared with separate download and review steps.

Content

Scanner excerpt · SKILL.md (reported line 31)May include surrounding context.

If this skill is not present in the current agent environment, install it with one command:

bash
curl -fsSL https://raw.githubusercontent.com/qomob/xclawskill/main/install.sh | bash

This installs the skill into the agent's skills directory and creates the xclaw-skill command. After installation, you may invoke the CLI either as xclaw-skill <action> or python3 scripts/xclaw_skill.py --action <action>.

Exfiltration Commands

High
Category
Prompt Injection
Confidence
90% confidence
Finding

The skill explicitly supports sending messages and broadcasting content to external agents over the network. In a skill context with shell, file, and state access, these communication primitives can be abused for data exfiltration if an agent is induced to transmit sensitive local information or credentials.

Content

Scanner excerpt · SKILL.md (reported line 59)May include surrounding context.

md
| "who am I" / "show my agent ID" / "check my identity" | `python3 scripts/xclaw_skill.py --action whoami --state-file /tmp/xclaw_state.json` |
| "send heartbeat" / "keep my agent online" | `python3 scripts/xclaw_skill.py --action heartbeat --state-file /tmp/xclaw_state.json` |
| "run as daemon" / "keep alive continuously" / "auto heartbeat" / "stay online" | `python3 scripts/xclaw_skill.py --action daemon --state-file /tmp/xclaw_state.json --interval 20` |
| "send message to agent" / "message agent <id>" / "tell agent <id>" | `python3 scripts/xclaw_skill.py --action send-message --state-file /tmp/xclaw_state.json --recipient-id "<uuid>" --content "<message>"` |
| "listen for messages" / "接收消息" / "stay connected and receive" | `python3 scripts/xclaw_skill.py --action listen --state-file /tmp/xclaw_state.json [--duration <seconds>]` |
| "broadcast to all agents" / "announce to network" | `python3 scripts/xclaw_skill.py --action broadcast --state-file /tmp/xclaw_state.json --content "<message>" --tags "<optional,filter>"` |
| "submit task result" / "task done" / "完成结果" | `python3 scripts/xclaw_skill.py --action submit-result --state-file /tmp/xclaw_state.json --task-id "<uuid>" --result '{"output":"..."}'` |

Chaining Abuse

High
Category
Tool Misuse
Confidence
96% confidence
Finding

Piping remote content directly into bash is a classic chaining-abuse pattern because it combines retrieval and execution into one opaque step. In the context of an agent-network skill that will attract users seeking automation, this is more dangerous because operators may normalize trusting unaudited installation flows, increasing the chance of full user-context code execution from any upstream compromise.

Content

Scanner excerpt · install.sh (reported line 11)May include surrounding context.

sh
# 同时安装 Python 依赖并创建 xclaw-skill 命令。
#
# 用法:
#   curl -fsSL https://raw.githubusercontent.com/qomob/xclawskill/main/install.sh | bash
#   或在仓库内直接运行: bash install.sh
#
set -euo pipefail

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

This skill implements self-upgrade that fetches remote git tags, checks out new code, and replaces the local working copy. That is a privileged capability unrelated to normal XClaw network operations, and it expands the attack surface from API interaction to arbitrary local code replacement if the repository, tag source, or local git remote is tampered with.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Executing git via subprocess for a network-interaction skill is an unjustified capability escalation because it lets the skill mutate local code and trust state. Even though arguments are hardcoded, the overall behavior enables remote-origin-driven code changes and therefore creates a supply-chain style execution path not needed for the advertised functionality.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The manifest enumerates network actions such as register, health, discover, send-message, broadcast, marketplace operations, balance, withdraw, reputation, semantic search, verify, and topology, but the README additionally promotes daemon, listen, setup, and self-upgrade commands. Of these, listen and daemon may support messaging presence, but setup and self-upgrade are clearly outside the manifest’s declared action set, creating a mismatch between claimed and documented capability.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

A self-upgrade capability lets the skill modify its own code or dependencies after installation, which expands trust from the reviewed package to future remote content. In an agent context, this is dangerous because users may approve a network interaction skill without realizing it can later fetch and apply code changes, potentially bypassing prior review.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
90% confidence
Finding

The skill documents substantial capabilities including shell execution, network access, filesystem writes, environment-variable use, package installation, and long-running processes, but it does not declare an explicit tool scope such as permissions or allowed-tools. This weakens policy enforcement and makes it harder for a host agent to constrain dangerous operations before execution.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The permissions table states that all credential sources are limited to explicit --api-key or the state file and explicitly says no credential-like environment variables are read. However, later documentation instructs users to supply the service endpoint through the XCLAW_BASE_URL environment variable, so the skill does read configuration from the environment despite the absolute claim. This is a documentation-level contradiction about environment-variable handling.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
77% confidence
Finding

The installer creates a persistent command in ~/.local/bin by symlinking xclaw-skill, which establishes a durable user-level execution path after installation. This is expected installer behavior, but persistence mechanisms are security-relevant because they leave executable artifacts that may later be invoked implicitly if the installed skill or target path is compromised.

Content

Scanner excerpt · install.sh (reported line 86)May include surrounding context.

sh
# ── 5. 创建 xclaw-skill 命令 ────────────────────────────────────────────
BIN_DIR="${HOME}/.local/bin"
mkdir -p "$BIN_DIR"
chmod +x "$TARGET/scripts/xclaw_skill.py"
chmod +x "$TARGET/scripts/xclaw-skill"
ln -sfn "$TARGET/scripts/xclaw-skill" "$BIN_DIR/xclaw-skill"

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The wrapper exposes additional actions such as setup, version, whoami, self-upgrade, daemon, and listen that are not declared in the skill manifest. Hidden or undocumented capabilities increase attack surface, can bypass user/operator expectations, and are especially risky here because some actions imply environment introspection, long-running listeners, or self-modifying behavior.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Administrative and runtime-control actions are exposed even though the stated skill purpose is primarily interacting with the XClaw network as a participant or observer. In this context, commands like self-upgrade, daemon, and listen could enable persistence, uncontrolled background execution, or execution flow changes that exceed the user's expected scope and create opportunities for abuse.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
94% confidence
Finding

The subprocess call itself is argument-safe because it does not invoke a shell, but it is part of a self-upgrade path that executes local git commands to fetch and switch code. That gives this skill an unnecessary local code-modification capability outside its stated XClaw interaction purpose, and if the configured git remote or fetched tag is malicious or compromised, the skill can replace its own code on disk.

Content

Scanner excerpt · scripts/xclaw_skill.py (reported line 1083)May include surrounding context.

python
import subprocess

    def _run(args):
        return subprocess.run(args, cwd=repo_root, capture_output=True, text=True)

    prev_head = _run(["git", "rev-parse", "HEAD"]).stdout.strip()
    tags = _run(["git", "ls-remote", "--tags", "--refs", "origin"])

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The listen loop prints incoming MESSAGE and BROADCAST payload content directly to stdout as JSON events. In many agent platforms stdout is centrally logged or exposed to calling systems, so received messages may be unintentionally disclosed to log readers, other tools, or retained telemetry without the operator realizing it.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest describes XClaw network participant and observer actions such as registration, messaging, marketplace operations, balance, reputation, topology, and verification, but it does not mention any self-upgrade or software update capability. This test explicitly invokes a self-upgrade action, indicating the skill code exposes an additional capability outside the stated purpose.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
85% confidence
Finding

The manifest describes the skill as a tool for interacting with the XClaw agent network through participant and observer actions, but the README also presents the skill as a shell-installed package fetched via remote install script. Installation behavior is not itself part of the stated network-operation scope, so the documented behavior exceeds the manifest’s functional description.

Content

No source excerpt is available for this finding.

External Script Fetching

Low
Category
Supply Chain
Confidence
74% confidence
Finding

This is the same underlying issue as the duplicate SC2 finding: remote script execution is being encouraged, and checksum verification from the same source only partially mitigates supply-chain risk. In an AI agent setting, explicit instructions to execute installers with user privileges materially raise the danger of compromise.

Content

Scanner excerpt · README.md (reported line 10)May include surrounding context.

推荐方式:下载后校验 SHA256 再运行(人与 AI Agent 均适用):

bash
curl -fsSL https://raw.githubusercontent.com/qomob/xclawskill/main/install.sh -o install.sh
shasum -a 256 install.sh          # 与仓库根目录 SHA256SUMS 对比
bash install.sh

External Script Fetching

Low
Category
Supply Chain
Confidence
74% confidence
Finding

This is the same underlying issue as the duplicate SC2 finding: remote script execution is being encouraged, and checksum verification from the same source only partially mitigates supply-chain risk. In an AI agent setting, explicit instructions to execute installers with user privileges materially raise the danger of compromise.

Content

Scanner excerpt · README.md (reported line 10)May include surrounding context.

推荐方式:下载后校验 SHA256 再运行(人与 AI Agent 均适用):

bash
curl -fsSL https://raw.githubusercontent.com/qomob/xclawskill/main/install.sh -o install.sh
shasum -a 256 install.sh          # 与仓库根目录 SHA256SUMS 对比
bash install.sh

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

The skill documentation switches substantial operational and safety guidance into Chinese, including the permissions table and several command examples, without stating that the user may choose their preferred language. This can create a locale/language policy issue because key instructions are effectively forced into a specific language for some users.

Content

No source excerpt is available for this finding.