Skip to content

Security: BaseIntelligence/agent-recipe

Security

docs/security.md

Security

Secrets

Secret Where it lives Notes
OPENROUTER_API_KEY Miner → Phala encrypted_env Required for live review/eval LLM; never commit or log
PHALA_CLOUD_API_KEY Miner host only Required for live CVM create; never bake into agent ZIP
Hotkey private material Wallet host only Remote-sign preferred; never in git or CI logs
Review session / one-time capabilities Challenge prepare responses CLI must not paste into tickets; treat as credentials

See .env.example for names only.

Gateway policy

  • Production scored path: no Base LLM gateway.
  • Vendored baseagent template calls refuse_if_gateway_residue fail-closed when Base gateway URL/token residue appears in the environment.
  • Use OpenRouter (or challenge-measured provider wiring) inside TEE only.

Image / ZIP integrity

  • Prefer challenge-documented measurement pins for review and eval images.
  • Do not ship unrelated secrets inside agent ZIPs.
  • Changing the agent template requires a new submit; attestation binds measured harness + agent bytes per challenge rules.

Money / spend safety

  • Prefer hard $20 cap and smallest CPU TDX shapes (tdx.small).
  • Refuse GPU targets on this product path (per agent-challenge selfdeploy guards).
  • Always teardown Phala inventory after runs.
  • Without PHALA_CLOUD_API_KEY, stay on dry-run / --help.

Residual risk

  • TEE isolation is strong but not absolute; see agent-challenge security docs.
  • Third-party Phala and OpenRouter availability/policy can block deploys.
  • Cryptographic attestation binds measurements and reports; it does not remove all operator or supply-chain risk.

Reporting

Do not open public issues that include keys, mnemonics, or prepare capabilities. Redact env dumps before sharing.

There aren't any published security advisories