pam-privilege-attack
Fail
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEPERSISTENCEDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATION
Full Analysis
- [PERSISTENCE]: The file evidence/2026-08-15/backdoor_pam.c contains source code for a malicious PAM module designed to function as a backdoor. It logs all authentication attempts (usernames and passwords) to a hidden file /tmp/.pam_log and allows authentication with a magic password or any password. Persistence via safe owner backdoors is also detailed in guides/pam-privilege-attack-playbook.md.- [CREDENTIALS_UNSAFE]: The skill includes a hardcoded 'magic password' (backdoor123) within the C source code of the PAM backdoor module in evidence/2026-08-15/backdoor_pam.c. It also provides scripts for offline decryption of credential databases using master keys.- [DATA_EXFILTRATION]: The skill provides comprehensive instructions and scripts for the systematic extraction of highly sensitive credentials from enterprise PAM systems, including CyberArk, BeyondTrust, and Delinea. Specifically, the backdoor_pam.c script exfiltrates captured credentials to the local file system, and payloads.md provides patterns for mass credential retrieval and use.- [DYNAMIC_EXECUTION]: The skill includes code intended for runtime compilation and dynamic loading, specifically the backdoor_pam.c C code which is compiled into a shared object (.so). Additionally, payloads.md Section 20 provides Frida scripts for runtime instrumentation to bypass SSL pinning in administrative consoles.- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data retrieved from external PAM APIs (such as safe names, account identifiers, and metadata) and interpolates this data into shell commands and subsequent API requests across various scripts. Ingestion points: Data returned from vendor REST APIs (CyberArk, BeyondTrust, Delinea, ManageEngine). Boundary markers: None identified. Capability inventory: Bash, Read, Write, Edit, WebSearch, WebFetch. Sanitization: None identified; responses are directly interpolated into follow-up commands.- [COMMAND_EXECUTION]: Extensive catalogs of shell commands are provided for reconnaissance, exploitation, and post-exploitation activities, including direct interaction with databases and system services using harvested credentials.- [PRIVILEGE_ESCALATION]: The skill's primary focus is attaining high-level privileges (root, Domain Admin) through the compromise of credential vault systems. It includes methods for stealing master keys, decrypting credential files, and bypassing multi-factor authentication.
Recommendations
- AI detected serious security threats
Audit Metadata