The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes multiple code snippets that demonstrate how to exfiltrate data from a target API to an external attacker-controlled domain.
Evidence in SKILL.md: Demonstrates exfiltration using fetch to https://attacker.com/log and navigator.sendBeacon to https://attacker.com/exfil.
Evidence in SCENARIOS.md: Demonstrates JSONP hijacking exfiltration to https://attacker.com/collect.
[DATA_EXFILTRATION]: The skill provides a playbook for performing reconnaissance on internal network addresses and the local loopback interface.
Evidence in SKILL.md: Includes a list of private IP ranges (192.168.1.1, 10.0.0.1, 172.16.0.1) and localhost services (Elasticsearch, Consul) to be probed via the victim's browser.
[COMMAND_EXECUTION]: The skill contains command-line instructions for manual testing of HTTP endpoints and CORS policies.
Evidence in SKILL.md: Provides curl commands to test for origin reflection and cache poisoning vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill defines a vulnerability surface where an agent processes untrusted data from external APIs and HTTP headers.
Ingestion points: SKILL.md (external API responses, Origin headers), SCENARIOS.md (JSONP callback payloads).
Boundary markers: The skill does not provide specific delimiters or instructions to ignore malicious commands embedded in the data it audits.
Capability inventory: The skill utilizes network request capabilities including fetch, curl, and navigator.sendBeacon to interact with and potentially exfiltrate data from the execution environment.
Sanitization: No sanitization logic is provided for the data ingested during the audit process.

cors-cross-origin-misconfiguration