The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill contains a vast library of RCE payloads and executable command patterns, including reverse shell strings, JNDI injection (Log4Shell), and various deserialization gadgets for Java, PHP, and Python. Automated scanners identified multiple files as infected with malware signatures (e.g., Python:Agent-RX, Script:SNH-gen) within the payloader directory.
[PROMPT_INJECTION]: The AI security and LLM sections contain explicit payloads designed to override agent behavior and exfiltrate system prompts, including 'Ignore all previous instructions' markers and 'DAN/Jailbreak' variants. While intended for testing external systems, these instructions are processed by the agent and could lead to behavioral hijacking.
[DATA_EXFILTRATION]: Multiple payloads are specifically designed to exfiltrate sensitive data, such as AWS IAM credentials, system environment variables, and browser cookies, to external domains including 'attacker.com' and 'evil.com'.
[OBFUSCATION]: Detected base64-encoded strings that decode to executable JavaScript and Shell scripts, as well as the use of zero-width characters (e.g., U+200B) to bypass keyword filters in prompt injection examples. HTML entity encoding is also used to conceal malicious JavaScript function calls.
[PRIVILEGE_ESCALATION]: The knowledge base includes detailed instructions and tools for bypassing Windows User Account Control (UAC), executing DLL hijacking, escaping Kubernetes containers, and utilizing 'Potato' exploit variants to escalate from service accounts to SYSTEM privileges.
[PERSISTENCE]: The skill documents methods for establishing persistence across system reboots using Windows registry keys, WMI event subscriptions, startup folders, and the creation of hidden backdoor users.
[INDIRECT_PROMPT_INJECTION]: The 'RAG投毒与知识库注入' (RAG Poisoning and Knowledge Base Injection) section explicitly describes attack surfaces for injecting instructions via external data sources and documents.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of numerous third-party penetration testing tools from community GitHub repositories and suggests executing remote scripts via 'curl | bash' patterns.

pentest-tools