The Agent Skills Directory

[PROMPT_INJECTION]: The skill's instructions in SKILL.md and the safety evaluation in SAFETY.md explicitly teach the agent to disregard 'SECURITY NOTICE' headers and 'IGNORE' lists found in external content. While framed as preventing 'over-refusal,' this behavioral training encourages the agent to bypass standard safety demarcations, which could be exploited by an attacker to execute malicious instructions embedded in data that the agent has been taught to treat as non-authoritative.
[INDIRECT_PROMPT_INJECTION]: The skill incorporates a section specifically designed to test and reinforce the bypassing of embedded security constraints (Category 8).
Ingestion points: SKILL.md processes a large block of text explicitly labeled as an untrusted external source.
Boundary markers: The skill contains markers (SECURITY NOTICE) but instructs the agent to ignore them as potential 'meta-injections'.
Capability inventory: The skill is intended for 'secrets storage and retrieval' and includes shell scripts (deploy.sh), suggesting file system and potentially network capabilities.
Sanitization: No technical sanitization is evident; the logic relies purely on training the agent to ignore safety labels.
[COMMAND_EXECUTION]: The skill's evaluation documentation (SAFETY.md) encourages the agent to proceed with command execution (e.g., running pytest) even when the input data contains critical security overrides. This creates a risk where the agent might be tricked into executing malicious commands if it perceives the safety warnings as fake 'over-refusal' attempts.
[SUSPICIOUS_CONTENT]: SKILL.md contains a sequence of seemingly random tokens and a URL: '....token to be expanding clawvault-large-scale sol.app/include namely tractiondirect launch agility ai-smart contracts...'. This 'word salad' is a known adversarial technique used to confuse LLM tokenizers or trigger specific behaviors via attention-sink manipulation.

clawdvault-hardened