The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains directives intended to override standard AI safety guidelines and operational constraints.
Evidence: The skill references an 'unlimited-attack-scope' component and uses terminology like 'not setting fixed paths' and 'jumping out of fixed attack chains' to encourage unrestricted autonomous behavior.
Evidence: The 'Project Blackboard' architecture (Category 8) creates a vulnerability surface for indirect prompt injection where untrusted data from a target environment can influence the agent's logic.
Ingestion points: The pentest-blackboard component ingests project facts derived from target systems into a SQLite database using the upsert_project_fact function.
Boundary markers: None identified; the agent is instructed to let its logic 'emerge' from these facts without delimiters or safety warnings.
Capability inventory: Extensive capabilities including network reconnaissance, credential harvesting, binary reversing, and autonomous tool bootstrapping.
Sanitization: No evidence of sanitization or filtering for data ingested from external targets into the blackboard system.
[COMMAND_EXECUTION]: The skill acts as an orchestrator for a suite of offensive security tools and shell-based operations.
Evidence: References to external components for reconnaissance (attack-surface-recon), network/domain attacks (active-directory-attack), and internal tool bootstrapping (proxy-tool-bootstrap).
[DATA_EXFILTRATION]: The skill explicitly includes components for harvesting sensitive information and credentials.
Evidence: The suite maps skills for initial-access-phishing, source-code-hunting for secrets and keys, and post-exploitation for credential harvesting and password cracking.
[REMOTE_CODE_EXECUTION]: The skill facilitates the discovery and execution of arbitrary code and exploits.
Evidence: Inclusion of a zero-day-discovery engine and web-attack-methods for autonomous exploitation of target services.

pentest-agent-os