The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses prescriptive instructions like 'CRITICAL: 路由执行契约（必须立即执行）' and an entire 'Obedience Engineering' framework in llm-security/references/agent-obedience-engineering.md designed to override model guardrails and autonomy.
[OBFUSCATION]: Filenames and internal links in seed-009_ios-jailbreak-detect-bypass.md use HTML character entities and percent encoding to mask keywords like 'jailbreak'. Payloads in field-journal/seed-012_log4shell-jndi-rce.md utilize Base64 encoding to hide executable commands containing remote URLs.
[EXTERNAL_DOWNLOADS]: The scripts/bootstrap-reverse.ps1 and .sh infrastructure fetches multiple third-party tools, plugins, and repositories from GitHub (e.g., skylot/jadx, NationalSecurityAgency/ghidra) and various package registries at runtime.
[REMOTE_CODE_EXECUTION]: The bootstrap system executes downloaded content and manages environment paths. Functions like Ensure-PipPackageInstall and Ensure-GitCloneInstall download and execute code from remote sources without local verification of the remote content beyond commit pinning.
[DATA_EXFILTRATION]: Educational payloads in field-journal/seed-017_xxe-oob-exfil.md demonstrate methods for exfiltrating local system files (e.g., /etc/passwd) to a remote attacker-controlled server using OOB techniques.
[INDIRECT_PROMPT_INJECTION]: The skill is specifically designed to ingest and process untrusted external files (malware, APKs, binaries).
Ingestion points: Untrusted binaries and APKs loaded for analysis in modules such as malware-analysis/ or apk-reverse/.
Boundary markers: The ops/scope-contract.md requires authorization status to be set to 'granted' before action, but the skill lacks strict delimiters for processing untrusted file content.
Capability inventory: The skill has access to full subprocess execution, file system modification, and network operations through its suite of scripts.
Sanitization: Minimal sanitization is performed on data extracted from external files before it is processed by the agent.

reverse-skill-router