The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core design of fetching, reading, and synthesizing untrusted web content.
Ingestion points: The skill uses web_search and web_fetch to ingest data from diverse and untrusted internet sources (Reddit, Twitter/X, and general websites) as described in references/research-protocol.md.
Boundary markers: While the skill uses a structured template (references/brain-template.md), it lacks explicit markers or instructions to isolate the retrieved data or prevent the agent from obeying instructions embedded within the fetched content.
Capability inventory: The skill possesses the ability to read and write files to the local brains/ directory, perform network operations via search/fetch tools, and potentially spawn additional agent processes for adversarial debates.
Sanitization: No sanitization or filtering mechanisms are defined for the content retrieved from external sources before it is synthesized into the agent's decision-making frameworks.
[EXTERNAL_DOWNLOADS]: The research protocol instructs the agent to perform broad and deep web fetches from untrusted third-party sources to build its knowledge base, creating a large attack surface for malicious data ingestion.
[COMMAND_EXECUTION]: The skill relies on frequent filesystem interactions, specifically reading and writing markdown files in the brains/ directory to manage its long-term knowledge storage.

deep-learning-agent