Feedback endpoint security: Phase 2 — registration + proof-of-work #3
Description
Context
Phase 1 security is implemented (rate limiting, agent fingerprinting, honeypot, input validation, dedup). This issue tracks Phase 2.
Phase 2 Security Measures
Instance Registration
- Agent registers with Portal on first run, receives a unique instance token
- Token sent with all subsequent requests (feedback, dispatch polling)
- Enables per-agent rate limiting and analytics
- Registration should be seamless — happens behind the scenes during
instar init
Lightweight Proof-of-Work
- Each feedback submission includes a small computational challenge solution
- Challenge issued by endpoint, agent computes and returns
- Raises cost of spam without burdening legitimate agents
- Could be simple hash prefix challenge (find nonce where SHA256(payload + nonce) starts with N zeros)
Content Fingerprinting
- Detect near-duplicate submissions across different agents
- Flag coordinated spam campaigns
- Use similarity hashing (simhash or minhash) on description field
Behavioral Analysis
- Track submission patterns per registered agent
- Flag anomalies (sudden burst, unusual types, templated descriptions)
- Auto-quarantine suspicious submissions for review
Design Principle
Security through intelligence, not just cryptography. The agent's grounding and coherence serve as the primary defense layer. Cryptographic measures are the foundation, but the agent's ability to evaluate and reason about what it's sending/receiving is the real security.
Related
- Intelligent self-update: npm update lifecycle management #1 (self-update lifecycle)
- Intelligence dispatch system: broadcast guidance to agents #2 (dispatch system — needs same security)
- Phase 1 implementation (rate limiting, fingerprinting, honeypot — already live)