cloud-native-vuln-research
Warn
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONOBFUSCATION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches tools and exploit code from numerous external sources. It downloads Syft from Anchore (Trusted Org) and k3d from its official site (Well-Known Service). However, it also instructs the agent to clone repositories and download PoCs from several unverified GitHub accounts, including feihong-cs, BobTheShoplifter, and Notselwyn, which represent an untrusted supply chain risk. It also queries security research platforms such as sploitus.com and vulners.com that are outside the whitelisted domains.
- [REMOTE_CODE_EXECUTION]: Several commands involve executing remote code or downloaded scripts. It utilizes piped shell patterns (curl | bash and curl | sh) for tool installation. It further directs the execution of downloaded exploit code using java -jar, python3, and binary execution after local compilation.
- [COMMAND_EXECUTION]: The skill makes heavy use of high-privilege system commands. This includes tool installation with sudo apt install and executing container operations with the --privileged flag and --cap-add=SYS_ADMIN capability.
- [CREDENTIALS_UNSAFE]: Instructions are provided to access and manipulate sensitive system credentials and tokens. This includes reading /etc/passwd and /etc/shadow files, as well as harvesting Kubernetes ServiceAccount tokens and cloud provider metadata (IMDS) from internal endpoints.
- [PRIVILEGE_ESCALATION]: The core content provides step-by-step guides for reproducing privilege escalation vulnerabilities. It specifically covers techniques to achieve host root access from within a container using kernel exploits such as CVE-2022-0185 and Dirty Pipe.
- [DYNAMIC_EXECUTION]: The skill involves creating and executing code at runtime. It uses gcc to compile C source code for kernel-level exploits and constructs dynamic JNDI injection payloads.
- [OBFUSCATION]: Exploit examples contain Base64-encoded command strings (e.g., dG91Y2ggL3RtcC9wd25lZAo=) which decodes to executable shell commands used to demonstrate code execution.
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface by ingesting data from external vulnerability APIs (NVD, FIRST, Sploitus).
- Ingestion points: Data is fetched from services.nvd.nist.gov and api.first.org via curl.
- Boundary markers: None observed; the skill assumes data retrieved from these APIs is reliable for processing.
- Capability inventory: The skill has extensive command execution and network capabilities across all referenced scripts.
- Sanitization: External data is processed using jq without explicit sanitization or validation steps.
Audit Metadata