The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements significant defenses against malicious instructions embedded in email data. The scripts/sanitize.py module aggressively filters HTML, Unicode control characters, and steganographic markers, wrapping the resulting text in a nonce-bearing fence. This structure ensures that an LLM treating the content as data cannot easily be coerced into executing injected commands. Test cases in tests/poisoned/ explicitly validate these mitigations against known attack vectors like delimiter breakout and envelope spoofing.\n- [DATA_EXPOSURE]: The skill manages access to sensitive email data by using a deterministic engine that operates locally. It minimizes data exposure to LLMs by truncating inputs and using a constrained JSON action space for model outputs, ensuring that the agent cannot be tricked into exfiltrating information or modifying sensitive configuration files like rules.yaml.\n- [UNVERIFIABLE_DEPENDENCIES]: The skill relies on standard, trusted libraries such as PyYAML and certifi. The installation process documented in scripts/install.sh uses standard package management and emphasizes integrity checks for the local engine runtime.\n- [DYNAMIC_EXECUTION]: The engine is designed with architectural separation, where the LLM only proposes changes to rules, while a deterministic Python script performs the actual execution. This prevents the dynamic execution of instructions found within email bodies.

synthesis-inbox-cleanup