The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'context stuffing' technique by including approximately 1,200 redundant sections of technical filler across eight reference files (e.g., architecture-patterns.md, state-management.md, testing-strategies.md). Each file contains 150 sections of rotated technical prose regarding memory management, idempotency, and Zero Trust. This excessive volume of technical noise is a known pattern used to overwhelm the agent's context window and potentially displace safety instructions or system prompts.
[DATA_EXFILTRATION]: Reference documentation in references/methodology-phases.md provides explicit command-line templates for exfiltrating local data to an external domain (attacker.com) via HTTP POST requests, DNS queries, and ICMP 'ping' packets.
[COMMAND_EXECUTION]: The reference files include numerous high-risk command snippets for acquiring administrative privileges in cloud environments (e.g., aws iam create-policy-version), bypassing container security via cgroup escapes, and harvesting credentials from system memory (e.g., Invoke-Mimikatz). While presented as documentation, these provide a direct surface for the agent to execute dangerous state-modifying actions.

penetration-testing