Back to skill

Security audit

Chat to Podcast

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly a disclosed conversation-to-blog publisher, but it reaches broadly into historical chat logs and can publicly publish derived content, so users should review it carefully before installing.

Install only if you are comfortable with the agent reading historical OpenClaw session logs and publishing AI-derived content to a public Halo blog. Prefer using it only on explicitly selected current-session content, review the full preview carefully, avoid shared-agent environments, and pin or preinstall the Markdown converter and Halo CLI instead of allowing runtime npm resolution.

Vulnerability Patterns
  • 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
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
Findings (4)

T08 · Insecure Dependencies

Error
Location
scripts/publish-to-halo.sh:34
Finding

Unpinned Packages May Be Downloaded and Executed at Runtime

Content
View full analysis
HTML if npx marked "$TMP_DIR/body.md" > "$TMP_DIR/body.html" 2>/dev/null; then HTML_MODE=true else echo "Warning: HTML conversion failed, falling back to Markdown import." HTML_MODE=false fi ``` The installation instructions similarly install the latest globally available Halo CLI release without a version or integrity constraint: ```bash npm install -g @halo-dev/cli ``` If `marked` is not already available locally, `npx` can resolve and download it from the configured npm registry before executing it. Consequently, the code that runs is not fully represented by the audited project and can change after review. A compromised package version, registry, lockless dependency chain, or npm configuration could execute arbitrary code with the privileges of the account running the skill. Redirecting standard error does not mitigate this risk; it only hides relevant installation or execution diagnostics. ### Attack Path 1. The operator invokes `publish-to-halo.sh` on a system where `marked` is not locally pinned and installed. 2. `npx` resolves `marked` using the current npm configuration and registry state. 3. An attacker compromises the resolved package, one of its dependencies, the configured registry, or the dependency resolution path. 4. The malicious package executes during the Markdown conversion step. 5. The payload inherits the publishing process's user privileges and may access the user's files, OpenClaw session data, Halo credentials, environment variables, and network connectivity. ### Impact Assessment Successful exploitation provides arb ...[truncated 428 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:30
Finding

Skill Instructions Permit Overbroad Access to Historical Conversation Logs

Content
View full analysis
/dev/null) echo "$ts $f" done | grep "2026-04-15" ``` ```bash for f in ~/.openclaw/agents/main/sessions/*.jsonl; do if python3 -m json.tool "$f" 2>/dev/null | grep -qi "Halo"; then echo "MATCH: $f" fi done | head -10 ``` The associated instruction is: ```text After locating the target session, extract the complete conversation record without limiting the number of entries. ``` This approach crosses the least-privilege boundary between the conversation selected for publication and every session readable by the agent account. Keyword and date matching are not authorization controls, and the instructions do not require per-session ownership validation, participant validation, or explicit confirmation before loading a historical session into model context. Although the skill later requires confirmation before publication and instructs the agent to redact secrets, those controls occur after the historical data has already been read and processed. Redaction also relies on model behavior rather than deterministic filtering. ### Attack Path 1. An attacker or unauthorized user who can interact with a shared agent invokes the skill and supplies a broad keyword, date range, or ambiguous historical-session description. 2. The agent enumerates all JSONL session files available to its operating-system account. 3. The keyword search identifies ses ...[truncated 971 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/publish-to-halo.sh:52
Finding

Full Article Content Is Exposed Through Process Command-Line Arguments

Content
View full analysis
/cmdline`, another local user or monitoring component may read the full article while the `halo` process is running. The content originates from conversation logs and may therefore contain private or sensitive information even when redaction is intended. This also unnecessarily places potentially long content in the operating system's argument vector, which is subject to platform-specific length limits. ### Attack Path 1. A user publishes an article containing private conversation-derived content. 2. The script launches `halo post create` with the complete HTML body in the `--content` argument. 3. A local attacker, diagnostic collector, process monitor, or telemetry agent observes the Halo process before it exits. 4. The observer reads the command line and recovers the article body. 5. The recovered content can be stored or exfiltrated independently of Halo publication permissions. Exploitation requires local process-observation capability, and visibility varies according to operating-system and `/proc` hardening settings. ### Impact Assessment The exposure is limited to the article being processed, but that article can contain complete historical conversation material. Confidentiality impact may ...[truncated 212 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Note
Location
scripts/publish-to-halo.sh:29
Finding

Temporary Conversation-Derived Files Are Not Reliably Removed on Failure

Content
View full analysis
"$TMP_DIR/body.md" # Try Markdown -> HTML if npx marked "$TMP_DIR/body.md" > "$TMP_DIR/body.html" 2>/dev/null; then HTML_MODE=true else echo "Warning: HTML conversion failed, falling back to Markdown import." HTML_MODE=false fi ``` Cleanup is performed only on explicit success paths with: ```bash rm -rf "$TMP_DIR" ``` Because the script enables `set -e` and does not install an `EXIT`, `INT`, `TERM`, or `HUP` trap, failures in later Halo, Python, JSON parsing, or update commands can terminate the script before cleanup. Interruption can have the same result. This leaves conversation-derived Markdown or HTML on disk under the system temporary directory. `mktemp -d` generally creates a directory with restrictive permissions, which limits immediate cross-user exposure, but abandoned files remain accessible to the same account, administrators, backups, forensic tools, or a later account compromise. ### Attack Path 1. The script creates `body.md` and potentially `body.html` under a temporary directory. 2. Publication is interrupted, or a later `halo`, Python, or parsing command returns a nonzero status. 3. `set -e` terminates the script before an explicit `rm -rf` statement is reached. 4. The temporary directory remains on disk. 5. A process running as the same user, an administrator, or a subsequent attacker with access to the account discovers and reads the abandoned content. ### Impact Assessment The impact is residual disclosure of the article body and any sensitive conversation material it contains ...[truncated 208 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (11)

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

声明描述的是一个“对话转 Podcast 文稿并确认后发布”的完整流程,核心能力应包括对聊天内容进行抽取、整理和生成文稿,以及预览确认。实际代码没有任何对话读取、文本分析、关键词/时间范围筛选、文稿生成或交互确认逻辑。它只是把一个已有的 Markdown 文件发布到 Halo:提取 title/slug、去 front matter、可选 Markdown 转 HTML、调用 halo CLI 创建或导入文章并公开发布。因此,虽然‘发布到 Halo’这一部分与声明的一小部分一致,但整体主用途与关键能力明显不匹配,属于实质性描述—行为不一致。

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
89% confidence
Finding

The skill instructs use of file reads and shell commands against local session logs and the Halo CLI, but it does not declare any explicit tool scope or allowed-tools boundary. That makes the effective privilege set ambiguous and increases the chance of overbroad file/system access when the skill is invoked.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The display name and all user-facing instructions are fixed in Chinese, and the skill does not state that users may choose another language for prompts, previews, or published output. This creates a natural-language policy concern because the skill appears to enforce a single locale without opt-in or documented justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger phrases are broad and overlap with ordinary conversation, making accidental invocation more likely. In this skill, accidental invocation is more dangerous because the workflow includes reading session history and preparing content for publication, which can cause unintended data processing or disclosure.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill directs the agent to mine current and historical session logs, search by keyword/date, and extract complete conversation records. This creates a significant data exposure pathway because sensitive information from unrelated or older sessions may be collected, transformed, and potentially published, especially given the later blog-posting workflow.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill explicitly instructs the agent to search current and historical session logs and extract complete conversation records, including from unrelated prior sessions. That exceeds the minimum data needed for many user requests and can expose sensitive content from other conversations without granular consent.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
76% confidence
Finding

The instructions say not to expose tool details, yet they also direct preservation of reasoning traces, uncertainty, and the user's inner thoughts in published output. Publishing inferred internal reasoning or sensitive user thought processes can disclose private contextual information that the user did not intend to make public.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
83% confidence
Finding

Using npx marked without a pinned version allows the environment to fetch whatever package version is current at runtime. This creates a supply-chain risk: a compromised or breaking upstream release could execute unintended code or alter publication behavior during skill execution.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
72% confidence
Finding

The skill persists conversation-derived content into an external blog system and also updates a local counter file, creating durable storage of potentially sensitive session material. This is risky in context because the workflow draws from broad session history and can publish content publicly, turning transient chat data into long-lived records.

Content

Scanner excerpt · SKILL.md (reported line 246)May include surrounding context.

  1. 创建文章
    bash
    halo post create \
      --profile blog-danke \
      --name <slug> \
      --title <标题> \
    

Rp1

Medium
Category
MCP Rug Pull
Confidence
92% confidence
Finding

The script invokes npx marked without pinning an exact package version, which can cause an unreviewed package version to be fetched and executed at runtime if marked is not already installed locally. In a publishing workflow that processes user-controlled content and may run with access to blog credentials or a privileged Halo profile, this creates a supply-chain execution risk that could lead to credential theft or unauthorized blog actions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

This shell script includes natural-language comments such as '从 front matter 提取 title 和 slug' and '去掉 front matter,保留正文', which impose a specific language context without any indication that the skill is intentionally region-specific or that users can choose their preferred locale. The policy explicitly calls out language or locale constraints as violations when they are forced without opt-in or justification.

Content

No source excerpt is available for this finding.