The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill outlines methodologies for exploiting trust boundaries where untrusted data enters the context of a multi-agent system.
Ingestion points: SKILL.md identifies A2A endpoints, shared memory/cache tools, and long-term memory as key vectors.
Boundary markers: Analysis indicates that inter-agent communication often lacks robust identity verification, relying on protocol handshakes (nonces) rather than authentication.
Capability inventory: The skill notes that sub-agents may have access to privileged tools (read_root_flag, admin_*) that can be triggered by instructions relayed from a less-privileged agent.
Sanitization: Highlights the risk of sub-agents executing instructions found in data without sanitization or provenance checks.
[DATA_EXFILTRATION]: A provided Python probe script in SKILL.md demonstrates how to perform network operations to test for inter-agent communication leaks. Evidence: The script uses urllib.request.urlopen to interact with external endpoints defined by placeholders (https://target/agentA/chat).
[OBFUSCATION]: SKILL.md discusses the technique of using Unicode tag characters (U+E0000–E007F) to hide malicious payloads from human review. No instances of these characters being used for malicious obfuscation were found within the provided files.
[COMMAND_EXECUTION]: The skill includes a self-contained Python script for security researchers to run. It uses standard library modules to simulate an attack and verify if an instruction propagates between agents.

hunt-agent-multi-agent