The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of user-supplied prompts within a tracing context, which constitutes a surface for indirect prompt injection.
Ingestion points: Arbitrary strings are passed to the model.trace(prompt) method as seen in SKILL.md and references/tutorials.md.
Boundary markers: The examples provided do not demonstrate the use of delimiters or specific instructions to the model to ignore potential injections within the analyzed text.
Capability inventory: The skill uses the nnsight library to record and execute computation graphs, including the capability to send these graphs to remote NDIF servers when remote=True is specified.
Sanitization: No evidence of input validation, filtering, or escaping of the prompt content is present in the documentation.
[EXTERNAL_DOWNLOADS]: The skill references and facilitates connection to the NDIF (National Deep Inference Facility) infrastructure at ndif.us for remote execution of model internals analysis. This is a core feature of the nnsight library for handling large-scale models.
[CREDENTIALS_UNSAFE]: The documentation provides appropriate guidance for managing sensitive API keys for the NDIF platform by suggesting the use of environment variables (NDIF_API_KEY) or dedicated configuration objects, rather than hardcoding credentials.

nnsight-remote-interpretability