The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and analyzing data from external, untrusted sources such as Reddit, community reports, and public issue trackers (documented in references/external.md and references/comparison.md). This creates an attack surface where malicious instructions embedded in external content could attempt to influence the agent's technical recommendations or local probes.
Ingestion points: The skill explicitly directs the agent to search for "issues, Reddit and other community reports" (SKILL.md) and "official docs/source/changelogs" (references/external.md).
Boundary markers: The instructions advise the agent to "distinguish documented capability, locally observed behavior, inference and unknown" (SKILL.md) and "Resolve material contradictions by source authority" (SKILL.md). While these provide methodological rigour, there are no specific technical boundary markers or instructions to treat external data as untrusted text rather than instructions.
Capability inventory: The skill permits the agent to "Run a bounded local probe" (references/comparison.md) and use a "minimal parser/compiler/runtime probe" (references/library-api.md) to verify findings. If an injection successfully influences the agent's logic, these capabilities could be misused.
Sanitization: There are no explicit instructions to sanitize, escape, or filter the content fetched from external URLs before processing it within the agent context.

research