Back to skill

Security audit

house-skills-kit

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because it reports unsafe file-writing paths, unreliable financial calculations, and privacy gaps; the evidence does not establish malicious intent, and one analysis was partial.

Consider avoiding installation until the renderer’s file-path handling is reviewed and the financial calculators’ reported inconsistencies are addressed. If you use the customer-facing workflows, avoid sending identifiable customer information to an AI service unless your organization has approved the processing and suitable privacy controls. The Skillspector scan was partial: 139 of 140 components were scanned, and one file had limited static analysis.

Vulnerability Patterns
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (33)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied commands implement repository publication, which is not disclosed in the declared installation, usage, generation, or selection purposes. The push can transmit committed repository contents to GitHub using available SSH credentials. The remote URL contains a username placeholder, so successful execution requires substitution and appropriate access; the snippet alone does not establish malicious intent or successful data transfer.

Content

No source excerpt is available for this finding.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · skills/calc-toolkit/buy_vs_invest.py (reported line 42)May include surrounding context.

python
try:
        from ruleengine import get_rules as _gr
        rules, srcmap = _gr("invest", city=city)
        return rules, srcmap
    except Exception:
        return {}, {}

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · skills/calc-toolkit/holding.py (reported line 37)May include surrounding context.

python
try:
        from ruleengine import get_rules as _gr
        rules, srcmap = _gr("holding", city=city)
        return rules, srcmap
    except Exception:
        return {}, {}

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · skills/calc-toolkit/loanterm.py (reported line 33)May include surrounding context.

python
try:
        from ruleengine import get_rules as _gr
        rules, _ = _gr("loanterm", city=city)
        return rules
    except Exception:
        return {}

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · skills/calc-toolkit/lpr.py (reported line 62)May include surrounding context.

python
try:
        from ruleengine import get_rules as _gr
        rules, srcmap = _gr("lpr")
        return rules, srcmap
    except Exception:
        return {}, {}

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · skills/calc-toolkit/spacing.py (reported line 36)May include surrounding context.

python
try:
        from ruleengine import get_rules as _gr
        rules, _ = _gr("sunshine", city=city)
        return rules
    except Exception:
        return {}

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · skills/calc-toolkit/tax.py (reported line 129)May include surrounding context.

python
def fee_get(key, default):
        """读取费用参数:优先层4顶层覆盖值,否则 fees.key,否则 default"""
        if key in rules:          # --set 顶层覆盖优先
            return rules[key]
        return fees.get(key, default)

    # 中介费

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% 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 · CONTRIBUTING.md (reported line 49)May include surrounding context.

md
## 代码约定

- **零第三方依赖**:`calc-toolkit` 与 `render.py` 只用 Python 标准库(Python 3.8+)
- 一个工具一件事;新增工具在 `skills/calc-toolkit/SKILL.md` 与主 README 的工具表里同步登记
- 规则数据与代码分离:政策类数字改 `skills/calc-toolkit/rules/*.yaml`,不改代码
- 所有涉路径的脚本必须限定在仓库目录内(参考 `install.sh` 与 `render.py` 的路径校验)

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% 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.en.md (reported line 106)May include surrounding context.

md
python3 skills/calc-toolkit/tax.py --price 1500000 --area 89 --city 昆明 --first 1 --held 6 --unique 1

# 2) Use a skill — each directory is a self-contained SKILL.md
cat skills/buyer-first-home-advisor/SKILL.md

# 3) Generate a white-labeled advisory skill from a brand config
python3 render.py --config examples/kunming/kunming.buy.yaml --out output/my-brand/

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The renderer's documented repository confinement guarantees contradict its actual file-access sequence and output validation.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The configuration is opened and parsed before its containment check, and the checked output base does not constrain the subsequently appended skill name. Configuration-controlled archetype and module paths also lack containment checks, so the claimed repository confinement does not hold for either all reads or final writes.

Content

No source excerpt is available for this finding.

Tainted flow: 'skill_path' from open (line 141, file read) → open (file write)

Medium
Category
Data Flow
Confidence
99% confidence
Finding

The output directory includes a skill name derived from unvalidated configuration fields, including brand.slug and skill.domain. Absolute paths or traversal components can escape the checked output root and overwrite a writable SKILL.md elsewhere, potentially poisoning another agent skill.

Content

Scanner excerpt · render.py (reported line 142)May include surrounding context.

python
os.makedirs(outdir, exist_ok=True)

    skill_path = os.path.join(outdir, "SKILL.md")
    with open(skill_path, "w", encoding="utf-8") as f:
        f.write(text)
    with open(os.path.join(outdir, "manifest.json"), "w", encoding="utf-8") as f:
        json.dump(manifest, f, ensure_ascii=False, indent=2)

Tainted flow: 'outdir' from open (line 138, file read) → open (file write)

Medium
Category
Data Flow
Confidence
99% confidence
Finding

The manifest write uses the same unchecked, configuration-derived output directory as the skill write. A crafted configuration can therefore create or overwrite manifest.json outside the intended repository boundary with the renderer's filesystem privileges.

Content

Scanner excerpt · render.py (reported line 144)May include surrounding context.

python
skill_path = os.path.join(outdir, "SKILL.md")
    with open(skill_path, "w", encoding="utf-8") as f:
        f.write(text)
    with open(os.path.join(outdir, "manifest.json"), "w", encoding="utf-8") as f:
        json.dump(manifest, f, ensure_ascii=False, indent=2)

    print(f"[ok] {skill_path}  (archetype={manifest['archetype']}, modules={manifest['modules']})")

Sensitive financial-data collection and documentary evidence retention lack privacy safeguards.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The workflow solicits credit, income, housing, and down-payment-source information and later requires documentary evidence without specifying consent, minimization, access restrictions, or retention limits. These are legitimate brokerage needs, but following the instructions with identifiable records can expose sensitive financial information through AI inputs or poorly controlled evidence storage.

Content

No source excerpt is available for this finding.

The reported rule sources do not provide the manifest's promised parameter provenance and can disagree with the values actually used.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The calculator discards the rule engine's source map and prints raw rule values as sources even when explicit command-line arguments supply different effective values. Silent fallback on rule-loading errors further obscures provenance, allowing financial results to appear backed by rules that were not actually used.

Content

No source excerpt is available for this finding.

The comparison does not maintain its promised equal cash flows when the holding period exceeds the mortgage term.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

For holding periods longer than the mortgage term, the investment branch continues accumulating mortgage-equivalent contributions for the entire horizon, while the buying branch receives no corresponding post-payoff savings. This breaks the stated equal-cash-flow comparison and can materially bias a high-value property decision against buying.

Content

No source excerpt is available for this finding.

The documented commission → customer rebate → team allocation workflow allocates team shares from gross commission rather than the post-rebate balance.

Medium
Category
Excessive Agency
Confidence
92% confidence
Finding

The rebate is calculated and displayed but never deducted from the commission pool used for team allocation. In the advertised sequential workflow, this can overstate distributable funds and company retention; gross-based allocation could be legitimate, but that contractual assumption is not made explicit.

Content

No source excerpt is available for this finding.

The provident-loan branch says there is no house-age constraint but enforces a hard-coded house-age-plus-term limit.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The provident-loan branch hard-codes a house-age-plus-term cap of 65 and includes it in the minimum-term calculation despite stating that this branch has no house-age constraint. This can incorrectly shorten or eliminate estimated eligibility, and the hard-coded cap bypasses the advertised rule customization.

Content

No source excerpt is available for this finding.

The documented original-loan input mode does not derive the remaining principal before comparing prepayment options.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

In original-loan mode, the program assigns the original principal directly to the remaining-principal variable while reducing the term by installments already paid. The existing remaining_balance helper is never called, so valid inputs for an older loan produce incorrect balances, payments, and prepayment comparisons that could mislead substantial financial decisions.

Content

No source excerpt is available for this finding.

Net-yield CLI options documented as percentages are applied as decimal fractions.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The CLI documents vacancy, tax, and maintenance inputs in percentage units, but applies supplied values directly as decimal fractions. Following the documented interface can therefore produce severely incorrect net yields and investment verdicts, including misleading positive results when both vacancy and tax factors become negative.

Content

No source excerpt is available for this finding.

The documented highest-priority user-template layer silently fails to apply ordinary saved template values.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The template loader extracts the requested domain before appending it to the merge chain, but the merge loop attempts to extract that same domain again. Ordinary domain-wrapped templates consequently fail to override defaults or city rules without warning, undermining customized financial calculations across dependent tools.

Content

No source excerpt is available for this finding.

Provident-fund recognition ignores the configurable recognition mode despite explicitly promising that it can be overridden.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The program reads fund_mode but never uses it in provident-fund recognition, despite explicitly advertising that it can be overridden. Users applying local policy settings still receive fixed-logic classifications, potentially resulting in incorrect eligibility or interest-rate advice.

Content

No source excerpt is available for this finding.

The documented ability to override all tax standards is contradicted by hard-coded tax thresholds and rates.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

Although the tool promises configurable tax standards, area thresholds, holding-period exemptions, and the differential income-tax rate remain hard-coded. Changes supplied through the advertised override mechanism cannot affect these decisions, potentially producing materially incorrect transaction-cost estimates when policy or applicable rules differ.

Content

No source excerpt is available for this finding.

Customer-information submission and telephone verification lack privacy and consent safeguards.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow requires channel partners to submit customer information and staff to verify customers by telephone, without defining permitted fields, authorization, or customer notice. This creates a privacy-control gap at an organizational data-sharing boundary, although the business purpose is legitimate and no malicious recipient is identified.

Content

No source excerpt is available for this finding.

Real customer conversations are requested without a privacy warning.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

The review mode explicitly invites users to paste real customer conversations into the AI without first removing personal information or confirming authorization. Real-estate conversations can contain contact details, household circumstances, and financial information, making this a concrete route for unnecessary disclosure to the model service.

Content

No source excerpt is available for this finding.