The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill directs the agent to read highly sensitive credentials from the environment, including ANTHROPIC_API_KEY, SLACK_BOT_TOKEN, HTB_PASS, and HTB_TOKEN using a custom env-reader.py tool. Accessing a platform API key like ANTHROPIC_API_KEY is a significant risk as it could be exfiltrated if the agent is compromised by external data.\n- [PRIVILEGE_ESCALATION]: The instructions require the agent to use sudo for several tasks, such as managing OpenVPN tunnels and modifying the system /etc/hosts file. Granting an AI agent root-level privileges to modify system configuration files significantly increases the potential impact of a security breach.\n- [DATA_EXFILTRATION]: The skill automates the transmission of detailed reports including attack narratives and technical findings to an external Slack channel using a bot token. This behavior could be exploited to exfiltrate sensitive data if the report content is influenced by malicious target data.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the HackTheBox platform (machine descriptions, challenge files) while maintaining broad capabilities like sudo access and network communication. This creates a surface for multi-step injection attacks.\n
Ingestion points: Machine descriptions, challenge files, and platform metadata fetched via Playwright or the HTB API (e.g., in SKILL.md and workflow.md).\n
Boundary markers: No delimiters or safety instructions were identified in the provided instructions to separate untrusted data from agent instructions.\n
Capability inventory: Root-level command execution (sudo), system file modification (/etc/hosts), network exfiltration via Slack, and autonomous browser control.\n
Sanitization: There is no evidence of input validation or sanitization for data retrieved from the HackTheBox platform.\n- [COMMAND_EXECUTION]: The skill makes extensive use of shell commands for reconnaissance and network interaction. It specifically instructs the agent to operate autonomously with a headed browser to bypass bot detection, reducing user oversight of the agent's actions.\n- [REMOTE_CODE_EXECUTION]: The workflow includes piping responses from HackTheBox's official API directly into Python for data parsing. This pattern is used for tasks such as identifying lab servers and verifying machine ownership status.

hackthebox