Skip to content

Security: abhijitkrm/cometcli

Security

SECURITY.md

Security Policy

cometcli manages validator infrastructure — we take that seriously.

Threat model

  • Consensus keys: priv_validator_key.json contents are never read into memory, logged, or sent to an LLM. priv_validator_state.json is read only for double-sign protection (HRS checks).
  • Ops keys: live in the OS keychain or an encrypted file backend. Never printed, never in LLM context — the redaction layer scrubs mnemonics, key material, and secrets in both directions.
  • Transactions: always simulated first, always displayed decoded, always require explicit human approval. No auto-broadcast path exists.
  • LLM traffic: only tool descriptions, redacted tool results, and a bounded status snapshot leave the machine. agent.provider = off disables it entirely.
  • Audit: append-only JSONL at ~/.cometcli/audit/ records every tool call, shell command, prompt (redacted), and tx.

Reporting

Report vulnerabilities privately via GitHub Security Advisories. Do not open public issues for exploitable findings.

There aren't any published security advisories