The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides proof-of-concept JavaScript and HTML code designed to exfiltrate data from a WebSocket connection to an external domain for testing purposes.
Evidence includes the use of fetch() and navigator.sendBeacon() in SKILL.md to send intercepted messages and sensitive profile data to https://attacker.com/collect, https://attacker.example.net, and https://attacker.com/data.
[EXTERNAL_DOWNLOADS]: The skill recommends installing third-party tools from external package registries.
Evidence includes instructions for pip install wsrepl and references to tools like wsdump from the websocket-client package and protobuf-inspector.
[COMMAND_EXECUTION]: The skill provides examples of executing command-line tools for security testing, fuzzing, and debugging.
Evidence includes instructions for running ws-harness.py, sqlmap for automated exploitation, and binary analysis tools like xxd and protoc.
[INDIRECT_PROMPT_INJECTION]: The skill describes a workflow involving the ingestion and analysis of untrusted data from external sources such as network traffic and JavaScript bundles, creating a potential attack surface.
Ingestion points: Intercepted HTTP/WebSocket traffic, binary frames (Protobuf/MessagePack), and JS bundles analyzed during testing.
Boundary markers: None explicitly defined in the skill instructions to protect the agent while processing this data.
Capability inventory: The skill instructions involve network operations (WebSockets), file system interaction (reading message files), and command execution for tooling.
Sanitization: While the skill discusses application-level sanitization for WebSockets, it does not define internal sanitization rules for the agent itself when handling untrusted input strings or binary data.

websocket-security