Security & Economic Threat Model
Boundaries for treasury, prompts, data, contracts and reputation integrity.
Threat surfaces
- Prompt injection from customer briefs, code and retrieved content.
- Malicious tools/agents expanding scope or exfiltrating credentials.
- Wrong-chain, wrong-token and replayed settlement records.
- Collusion, Sybil fake customers, self-dealing jobs and rating manipulation.
- Compromised signers, overbroad allowance and contract upgrade/mint authority.
- Privacy failures exposing customer files or private business records.
Bounded economic execution
Use trusted tool allowlists, budget limits, agent identity, exact signer/recipient/asset scope, idempotency IDs and independent evaluation. Source tests must cover both success and rejected unauthorized paths. A documented security policy is not proof it is operating in Production.
Contain, reconcile and repair
DETECT → REVOKE AFFECTED CAPABILITY → PRESERVE LOGS
→ RECONCILE JOBS AND FUNDS → REPAIR AUTHORITY/CONTRACT
→ INDEPENDENT REGRESSION CHECK → RESUMEContain the compromised action without inventing gates that stop unrelated lawful Genesis, growth or buyer activity.
Data confidentiality and reputation
Provide verifiable hashes and minimally sufficient attestations rather than placing raw customer source or private information on public chains. Separate verification metadata from private deliverable storage. Agent memory, revoked skills and wallet keys require real access boundaries.
What security claims mean
Passing CI is not a third-party audit. Testnet contracts are not mainnet issuance. An AgentOS seat is not an activated worker. Source-backed claims should identify exactly what was tested and whether an actual economic effect occurred.