aginti-agentlink

Pass

Audited by Gen Agent Trust Hub on Oct 19, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves the agent ingesting and summarizing external data from peer agents, which could contain malicious instructions designed to influence the current agent's behavior.\n
  • Ingestion points: Peer-provided AGENTS.md files, handoff documents, and .jsonl session history mirrors referenced in SKILL.md and references/protocol.md.\n
  • Boundary markers: The skill encourages the use of 'curated handoff docs' to separate operational facts from raw history, acting as a manual filter, though it lacks strict instruction delimiters or sanitization protocols.\n
  • Capability inventory: The agent has capabilities to execute local Python scripts (agentlink_handoff.py, agentlink_jsonl_summary.py), run remote commands via SSH, and write to the local filesystem.\n
  • Sanitization: The provided scripts/agentlink_jsonl_summary.py script truncates long lines but does not filter or sanitize the actual content of the agent history for potential instructions.\n- [DATA_EXFILTRATION]: The skill instructs the agent to perform extensive discovery and documentation of the local network and system environment, which could be sensitive if leaked.\n
  • Evidence: SKILL.md requires agents to build a 'Peer Map' containing hostnames, IP addresses, usernames, and absolute repository paths of all collaborating nodes.\n
  • Evidence: references/protocol.md describes a process for 'mirroring' history files from remote hosts using SSH and Base64 encoding to move raw session data between machines.\n- [COMMAND_EXECUTION]: The protocol explicitly authorizes the agent to execute shell commands and control hardware based on 'Action Contracts'.\n
  • Evidence: references/protocol.md provides examples of using ssh.exe for remote command execution and arduino-cli for local hardware control as part of the collaboration workflow.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 19, 2026, 06:25 PM