The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill accesses several environment variables containing sensitive API keys: ETHERSCAN_API_KEY, BLOCKSCOUT_API_KEY, BUNGEE_API_KEY, SOCKET_API_KEY, and LAYERSWAP_API_KEY. While SKILL.md contains instructions for the agent to avoid leaking these values, the skill's operational model depends on the exposure of these secrets to the runtime environment.
[EXTERNAL_DOWNLOADS]: The skill performs extensive network operations to fetch data from various blockchain explorers and bridge services, including etherscan.io, blockscout.com, debank.com, across.to, bungee.exchange, debridge.finance, gas.zip, hop.exchange, layerswap.io, layerzero-api.com, li.quest, relay.link, and symbiosis.finance. Several referenced RPC endpoints, including robinhoodchain.blockscout.com, fraxtal.gateway.tenderly.co, and polygon-bor-rpc.publicnode.com, have been flagged by automated scanners for having a poor reputation or being associated with phishing. Consequently, several configuration files containing these URLs (e.g., references/generated/target-mainnets.json and references/generated/target-fallback-rpcs.json) were also flagged as malicious by file reputation scanners.
[DYNAMIC_EXECUTION]: The file scripts/debank-collect.js provides JavaScript code that the agent is instructed to inject into a chromium-browser instance using evaluate_script. This code monkey-patches window.fetch to intercept and record API responses from DeBank, representing dynamic code execution in a web browser context.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests data from a vast array of external sources, including API responses and web page snapshots, which are then processed by the agent to satisfy user requests.
Boundary markers: The instructions emphasize a "strict read-only boundary" and separating facts from inference, but the skill lacks technical delimiters to separate untrusted external data from system instructions.
Capability inventory: The skill can execute local scripts, use network tools (curl), and control a browser, providing a significant set of actions that could be targeted by a malicious payload embedded in blockchain data.
Sanitization: The skill relies on jq and scripts/sweep-core.py to parse and validate incoming data, which helps mitigate some risks associated with malformed structured data.
[COMMAND_EXECUTION]: The skill executes various local scripts (chain-lookup.sh, blockscout-detect-plan.sh, etherscan-detect-plan.sh, debank-gate.py, and sweep-core.py) and interacts with external CLI tools (routemesh, etherscan, chromium-browser, and cli-cast). It uses uv run to execute Python logic, which implies a reliance on local command-line tools for its core operations.

evm-atlas