> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ameo.agiwithai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Security

> Threat model, key custody, and failure modes.

## Threat model

| Threat                       | Mitigation                                                                |
| ---------------------------- | ------------------------------------------------------------------------- |
| Unauthorized live execution  | `LIVE_ENABLED`, `WORKER_MODE`, operator-only endpoints                    |
| Model bypassing risk limits  | Deterministic [policy predicates](/glossary#policy-predicate) outside LLM |
| Key exfiltration via prompts | Private keys never sent to LLM providers                                  |
| RPC manipulation / outage    | Observation quality gate; honest error states in UI                       |
| Runaway trading volume       | `MAX_DAILY_VOLUME_USD`, per-trade caps                                    |

## Key custody

**Development / testnet:** AMEO uses a hot EOA (`AGENT_PRIVATE_KEY`) for signing. This is acceptable only for Sepolia demos with minimal funds.

**Production recommendation:** Replace hot EOAs with KMS, MPC, or a hardware-backed signing service. Never commit keys; use host secret managers.

## Policy defaults

* Drawdown floor, asset whitelist, max trade USD, slippage bps — see [Policy specification](/policy-spec).
* Failed predicates block execution and log `violations[]`.

## RPC failure

If Mantle RPC is unavailable:

* Observation quality drops; non-`no_op` plans may fail `degraded_observation_quality`.
* Worker returns 503-class errors; UI shows honest empty/error states.

## LLM provider outage

Fallback order: configured provider chain → `local_rules` deterministic planner.

When all remote providers fail, AMEO can still refuse unsafe actions — it does not silently disable guardrails.

## Contract verification

Run Slither / Foundry tests before mainnet deployment. Verify source on Mantlescan for public auditability.

## Responsible disclosure

Report security issues to **[security@agiwithai.com](mailto:security@agiwithai.com)** (replace with your operational inbox before production launch). Do not open public issues for key-handling vulnerabilities.