The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions and templates in references/tooling-install.md for installing numerous security and reconnaissance tools from external GitHub repositories and package registries (e.g., ProjectDiscovery, OWASP, and various individual developers). While these tools are widely used in the security industry, downloading and executing third-party software at runtime introduces potential supply-chain risks.
[COMMAND_EXECUTION]: The skill relies extensively on shell command execution, providing hundreds of one-liners using curl, dig, nslookup, and other utilities to probe target infrastructure, as seen in references/probes-and-wordlists.md and references/recon-stack.md.
[PRIVILEGE_ESCALATION]: Installation and execution instructions for specific tools, such as masscan in references/tooling-install.md, suggest the use of sudo to acquire elevated system privileges for raw socket operations.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze data from untrusted target systems, which introduces a surface for indirect prompt injection.
Ingestion points: Content is retrieved from external domains and target applications via curl and dig commands (e.g., in references/probes-and-wordlists.md).
Boundary markers: The instructions do not employ explicit delimiters or warnings to the agent to distinguish external data from internal instructions when processing content into the prompt context.
Capability inventory: The skill environment has access to powerful utilities including curl, python, go install, and pip install (references/tooling-install.md).
Sanitization: Untrusted content is primarily filtered using utilities like grep and jq for technical extraction, which does not provide mitigation against prompt injection attacks.
[SAFE]: The skill incorporates robust defensive practices, including a mandatory authorization confirmation gate, guidance on data minimization, and explicit warnings against sharing sensitive PII or credentials with cloud-based models.
[SAFE]: Automated reputation scans flagged two domains (socketscan.io and whopostedwhat.com) as malicious or blacklisted. These are documented within the skill's reference files (references/specialized-osint.md and references/people-osint.md) as legitimate OSINT tools for blockchain and social media research respectively. Two reference files were also flagged by scanners, likely due to the high density of links to offensive security resources.

offensive-osint