The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches public blockchain evidence from reputable industry services such as Etherscan, Blockscout, and various bridge protocols (Across, Bungee, deBridge, etc.). These network interactions are strictly read-only and target well-known, established domains.
[COMMAND_EXECUTION]: The skill executes local Bash and Python scripts, such as sweep-core.py and resolve-chain.sh, to process chain metadata and manage API rate limits. These scripts are transparently provided within the skill, use standard libraries, and do not incorporate untrusted inputs into shell commands.
[CREDENTIALS_UNSAFE]: Robust instructions are in place to prevent the accidental exposure of API keys (e.g., ETHERSCAN_API_KEY, BLOCKSCOUT_API_KEY). The skill mandates value-free presence checks and explicitly forbids echoing or logging secrets in output transcripts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted blockchain data (transactions, logs) but applies deterministic validation via regular expressions and fixed predicates in the sweep-core.py script. The strictly defined read-only boundary prevents this data from influencing state-changing operations.
[METADATA_POISONING]: Automated scanner detections for references/atlas-overlays.json and references/generated/blockscout-chains.md were manually investigated. The files contain legitimate, non-executable blockchain configuration data and URL registries; the alerts are identified as false positives resulting from high URL density.

evm-atlas