The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill designates https://docs.arcium.com/developers as an official documentation source. Automated scanners have flagged this URL as a phishing site, which creates a significant attack surface for indirect prompt injection. If the site is compromised or malicious, it could provide instructions designed to exploit the agent's capabilities.
Ingestion points: SKILL.md (Documentation Sources section).
Boundary markers: Absent; the agent is instructed to use these as authoritative sources.
Capability inventory: solana-keygen (key management), wget (remote fetching), cargo install (package installation), and anchor build (arbitrary code execution).
Sanitization: No sanitization or validation of external documentation content is suggested.
[EXTERNAL_DOWNLOADS]: The environment setup guidelines (ENVIRONMENT.md) instruct the agent to download and install various developer tools from remote sources. These include:
Downloading platform-tools from the Anza GitHub releases repository (github.com/anza-xyz/platform-tools).
Installing the quasar-cli directly from a specific git revision on GitHub (github.com/blueshift-gg/quasar).
Installing the anchor-cli from the official crates.io registry.
[METADATA_POISONING]: The skill uses a consistent fictional or future-dated context, referencing technology states and regulatory documents from 2026 (e.g., an SEC interpretation 33-11412.pdf and Switchboard deprecations). This misleading metadata could cause the agent to misinterpret the current state of technology or regulatory compliance.

solana