T05 · Unauthorized Access and Privilege Escalation
- Location
locales/en/SKILL.md:195- Finding
Project-Controlled Switch Can Bypass Authorization and Backup Requirements for Destructive Operations
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent authorized web-testing rule set, but it includes documented bypass switches that can weaken authorization, destructive-action, and third-party-scope controls.
Review before installing. Use this only for clearly authorized testing, prefer signed `AUTHORIZATION.json`, avoid `auth.disable` and `MOWEN_AUTH_OVERRIDE` bypasses, do not allow destructive or third-party tests without separate written scope, and keep reports in a private authorized location with sensitive data redacted.
locales/en/SKILL.md:195Project-Controlled Switch Can Bypass Authorization and Backup Requirements for Destructive Operations
locales/en/SKILL.md:205Local Marker Substitutes for Independent Authorization to Test External Third-Party Systems
locales/en/SKILL.md:371SSRF Guidance Explicitly Encourages Cloud Metadata and Internal Protocol Probing
locales/en/SKILL.md:81Authorization Verification Can Be Disabled Through Untrusted Project or Environment State
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
rm -rf /tmp/web-security-test-rules
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
rm -rf /tmp/web-security-test-rules
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
rm -rf /tmp/web-security-test-rules
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
rm -rf /tmp/web-security-test-rules
The README defines intentionally obfuscated 'escape hatch' values, including one that explicitly skips authorization verification. Hidden bypass tokens materially undermine the claimed safety model and make unauthorized security testing easier, especially because the surrounding skill context is web security testing where authorization is the primary control.
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
# 卸载当前项目中的 Skill
rm -rf .codebuddy/skills/web-security-test-rules
# 卸载全局 Skill
rm -rf ~/.codebuddy/skills/web-security-test-rules
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
rm -rf .codebuddy/skills/web-security-test-rules
rm -rf ~/.codebuddy/skills/web-security-test-rules
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
rm -rf .codebuddy/skills/web-security-test-rules
rm -rf ~/.codebuddy/skills/web-security-test-rules
Rule 22 is internally inconsistent: earlier text says mowenbrokentrue cannot bypass backups, but the later clause says it can skip the mandatory 'explicit authorization + backup first' requirement. In an adversarial or rushed operational context, the weaker interpretation could be used to justify destructive actions without recoverable backups, creating real risk of irreversible data loss.
The skill introduces an escape hatch that weakens destructive-operation safeguards by allowing mowenbrokentrue to bypass explicit destructive authorization, while adjacent text elsewhere says those controls are mandatory and non-waivable. In a security-testing skill, this contradiction is dangerous because an agent or operator may rely on the weaker branch and perform destructive actions without clear authorization, increasing the risk of unauthorized data loss or service impact.
Rule 22 first states backup is mandatory and not waivable, but later redefines mowenbrokentrue as enabling a 'destructive one-click' mode that skips the same gate. This creates a policy inconsistency at the most sensitive point in the skill and can directly normalize or authorize high-risk destructive behavior under unclear conditions.
Code accesses a cloud instance metadata endpoint (e.g. 169.254.169.254). A single request can return temporary IAM credentials, making this a high-value SSRF target for credential theft.
- **D3 Business logic / concurrency**: race conditions (concurrent coupon / transfer / order), negative / over-limit, duplicate submission, flow bypass (skip validation steps).
- **D4 Privilege-escalation combos**: vertical + horizontal escalation combos, IDOR bulk traversal, enumerable UUID / sequence, function-level escalation (change role param).
- **D5 Injection extensions**: command injection, XXE, SSTI, deserialization, LDAP / NoSQL injection (not limited to SQL / XSS).
- **D6 SSRF**: internal / metadata probing (e.g. `169.254.169.254`), DNS rebinding, protocol tunneling (gopher / file); probe only, don't read large sensitive data. **SSRF PoC must not point to external third-party systems outside authorized scope — must not use the target server as an attack springboard to hit off-site / real-internal third parties**.
- **D7 File upload**: type / content bypass (webshell), path traversal, decompression bomb, overwrite existing files; use harmless samples, delete after testing.
- **D8 Info leak**: `.git` / `.env` / backup exposure, unauthorized Swagger / Docs, verbose errors / stack traces, directory listing.
- **D9 Config & transport**: missing security headers (CSP / HSTS / X-Frame-Options), Cookie attributes (Secure / HttpOnly / SameSite), weak TLS config, CORS misconfig.
Code accesses a cloud instance metadata endpoint (e.g. 169.254.169.254). A single request can return temporary IAM credentials, making this a high-value SSRF target for credential theft.
- **D3 Business logic / concurrency**: race conditions (concurrent coupon / transfer / order), negative / over-limit, duplicate submission, flow bypass (skip validation steps).
- **D4 Privilege-escalation combos**: vertical + horizontal escalation combos, IDOR bulk traversal, enumerable UUID / sequence, function-level escalation (change role param).
- **D5 Injection extensions**: command injection, XXE, SSTI, deserialization, LDAP / NoSQL injection (not limited to SQL / XSS).
- **D6 SSRF**: internal / metadata probing (e.g. `169.254.169.254`), DNS rebinding, protocol tunneling (gopher / file); probe only, don't read large sensitive data. **SSRF PoC must not point to external third-party systems outside authorized scope — must not use the target server as an attack springboard to hit off-site / real-internal third parties**.
- **D7 File upload**: type / content bypass (webshell), path traversal, decompression bomb, overwrite existing files; use harmless samples, delete after testing.
- **D8 Info leak**: `.git` / `.env` / backup exposure, unauthorized Swagger / Docs, verbose errors / stack traces, directory listing.
- **D9 Config & transport**: missing security headers (CSP / HSTS / X-Frame-Options), Cookie attributes (Secure / HttpOnly / SameSite), weak TLS config, CORS misconfig.
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
- **D4 Privilege-escalation combos**: vertical + horizontal escalation combos, IDOR bulk traversal, enumerable UUID / sequence, function-level escalation (change role param).
- **D5 Injection extensions**: command injection, XXE, SSTI, deserialization, LDAP / NoSQL injection (not limited to SQL / XSS).
- **D6 SSRF**: internal / metadata probing (e.g. `169.254.169.254`), DNS rebinding, protocol tunneling (gopher / file); probe only, don't read large sensitive data. **SSRF PoC must not point to external third-party systems outside authorized scope — must not use the target server as an attack springboard to hit off-site / real-internal third parties**.
- **D7 File upload**: type / content bypass (webshell), path traversal, decompression bomb, overwrite existing files; use harmless samples, delete after testing.
- **D8 Info leak**: `.git` / `.env` / backup exposure, unauthorized Swagger / Docs, verbose errors / stack traces, directory listing.
- **D9 Config & transport**: missing security headers (CSP / HSTS / X-Frame-Options), Cookie attributes (Secure / HttpOnly / SameSite), weak TLS config, CORS misconfig.
- **D10 Supply chain / components**: component fingerprinting + known CVE, debug endpoint exposure (actuator / phpinfo), dependency vulnerabilities.
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
"destructive": {
"allowed": false,
"items": [
"DELETE /api/orders/*",
"DROP TABLE temp_test"
],
"backup_required": true,
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
"destructive": {
"allowed": false,
"items": [
"DELETE /api/orders/*",
"DROP TABLE temp_test"
],
"backup_required": true,
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
### 破坏性操作备份记录
- 操作:<DELETE /api/orders/12345>
- 影响资源:<订单表 order_id=12345>
- 备份方式:<数据库导出>
- 备份位置:</backups/2026-08-15/orders_12345.sql>
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
### 破坏性操作备份记录
- 操作:<DELETE /api/orders/12345>
- 影响资源:<订单表 order_id=12345>
- 备份方式:<数据库导出>
- 备份位置:</backups/2026-08-15/orders_12345.sql>
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
mkdir -p .codebuddy/skills/web-security-test-rules
cp -r /tmp/web-security-test-rules/. .codebuddy/skills/web-security-test-rules/
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
mkdir -p .codebuddy/skills/web-security-test-rules
cp -r /tmp/web-security-test-rules/. .codebuddy/skills/web-security-test-rules/
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.
> 💡 说明:
> - `cp -r 源/. 目标/` 末尾的 `/.` 表示复制目录的**全部内容**(含隐藏文件),并避免多嵌套一层目录。
> - Windows 用户若无 `git` / `cp`,可用 Git Bash;或下载仓库 zip 解压后,将仓库内容整体放入 `.codebuddy/skills/web-security-test-rules/`。
> - 安装后的目录结构应为:`.codebuddy/skills/web-security-test-rules/SKILL.md`,CodeBuddy 据此识别该 Skill。
---
The activation keywords are broad terms like '安全测试', '渗透测试', and '隐蔽测试', which can trigger in ordinary discussion rather than an intentional invocation. In an agent skill that may alter behavior around security testing, overbroad triggers increase the chance of accidental activation and unsafe workflow changes.
The README claims the skill enforces signed authorization checks and generates reports, but this file provides only documentation and no verifiable implementation of those controls. In a security-testing skill, overstating safety gates can mislead users into believing unauthorized or unsafe behavior is technically prevented when it may not be.
The skill explicitly instructs writing reports and error logs to the filesystem, but it does not declare an allowlist such as permissions or allowed-tools. That mismatch weakens least-privilege controls: an agent may obtain broader file-write behavior than a reviewer expects, increasing the chance of overwriting files or persisting sensitive security-test artifacts in unintended locations.
The skill mandates generating, storing, sharing, and archiving security test reports that may include requests, responses, findings, and operational details, but it does not put a strong top-level privacy warning or strict minimization controls around that data handling. Because penetration-test artifacts often contain credentials, internal URLs, PII, or proof-of-access data, this creates a meaningful confidentiality and retention risk if reports are overshared or stored insecurely.