🤖 Agent Behaviour: Flagging Suspicious Feeds

This document defines how responsible agents should behave when encountering a .llmfeed.json that appears invalid, misleading, unsafe, or untrustworthy.

Flagging is part of the ecosystem's decentralized safety model.

✅ Why flagging matters

Agents may be autonomous — but trust must be earned.

Flagging enables:

  • 🛡️ Defense against abuse or hallucinated services
  • 📣 User transparency and warnings
  • 🧭 Ecosystem reputation tracking
  • 🔁 Feedback loops (LLMCA audits, future blocks)

🔁 Forms of flagging

Type Description
flags[] in the feed Embedded warnings (if the author accepts reputation tags)
flag.llmfeed.json Separate feed published by an auditor, LLM, user
submit to llmca.org Via API or proxy interface (preview, verify, Forge)

✅ Agent duties when flagging

Situation Action
Feed lacks trust or behaves strangely Warn user
Declared capabilities do not match actual Create flag entry
Signature mismatch Report immediately
User feedback negative (opt-in) Suggest flag or review

📄 Example flags[] block (inline or mirrored)

"flags": [
  {
    "type": "risk",
    "submitted_by": "agent://previewbot",
    "reason": "Declared `POST /run` does not exist",
    "status": "pending",
    "date": "2025-05-19T12:00:00Z",
    "source": "https://llmca.org/flags/234"
  }
]

🧩 Certification and escalation

If a certified feed receives repeated flags:

  • LLMCA may revoke the certification
  • Agents should fetch updated certification before trusting

📡 External flag feed (flag.llmfeed.json)

Auditors or automated agents may publish:

{
  "feed_type": "flag",
  "target": "https://example.org/.well-known/mcp.llmfeed.json",
  "reason": "Signed blocks do not match trust",
  "submitted_by": "https://trustbot.ai",
  "created_at": "2025-05-18T22:00:00Z"
}

🌐 Integration with LLMCA

  • Central flag database
  • Manual and automated flag review
  • Certification history tracking
  • Public reputation layer

📎 Related