physical-security-testing
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDATA_EXFILTRATIONPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches documentation templates and security toolkits from well-known repositories, including TrustedSec's legal templates and Hak5's USB-Rubber-Ducky tools on GitHub. These downloads are standard in the security community for reconnaissance and payload generation.
- [COMMAND_EXECUTION]: Utilizes a wide range of shell commands for RFID/NFC manipulation (Proxmark3), network auditing (nmap, tcpdump), and hardware imaging. It includes sensitive commands such as
ddfor disk imaging andshredfor secure data deletion, which are intended for managing assessment artifacts and wiping implants. - [PRIVILEGE_ESCALATION]: Includes instructions for using
sudoto install software packages and interact with low-level system devices. This is necessary for configuring specialized peripheral hardware like the Proxmark3 and managing network interfaces. - [DATA_EXFILTRATION]: Provides specific example payloads for target exfiltration, such as PowerShell commands for DNS-based harvesting of system hostnames and usernames, and the establishment of reverse TCP shells.
- [PERSISTENCE]: Outlines persistent access techniques via reverse SSH tunnels configured through
autosshand systemd units on network implants, ensuring connectivity is maintained throughout multi-day engagements. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted environmental data—such as building reconnaissance notes and observed guard patterns—to build engagement timelines and attack plans. This represents a potential surface for indirect injection if the source physical data or observed logs are manipulated to include adversarial instructions.
- Ingestion points: External reconnaissance data entered into
recon_pack.mdandengagement_timeline.mdacross multiple skill files. - Boundary markers: No delimiters or warnings are present to isolate the processed target observation data from the agent's instructional logic.
- Capability inventory: Possesses powerful shell execution (
Bash), file manipulation (Write,Edit), and network fetch (WebFetch) tools. - Sanitization: No sanitization or validation routines were observed for the ingestion of physical-world observations or external site content.
Audit Metadata