Security and Risk
Ghast AI addresses commercial accountability for a defined service outcome. Secure execution also depends on data protection, runtime controls, payment infrastructure, and financing risk management.
Risk boundaries
| Risk | Design response | Boundary |
|---|---|---|
| Vague promise | Claim and evidence requirements | Only narrow, testable claims enter Verification |
| Failed or partial delivery | Verification, payment hold, and challenge | Evidence can be incomplete or ambiguous |
| Weak Creator incentives | Creator Bond and breach slash | Economic accountability aligns incentives but does not guarantee delivery |
| Frivolous challenge | Challenge Bond and duplicate-challenge controls | Economic friction reduces but cannot eliminate coordinated abuse |
| Reviewer error or collusion | Evidence-based review, conflict rules, and appeals | Conclusions remain bounded by the available evidence |
| Workflow or data leakage | Selective disclosure and access controls | Runtime permissions and data security remain essential |
| Compromised agent or tool | Runtime permissions and security | Protocol covers only properties in the claim |
| Non-deterministic output | Bounded criteria and tolerances | Objective verification varies by service |
| Contract or custody failure | Contract and custody controls | Technical and custody risk cannot be eliminated |
| Financing loss | Capability eligibility, senior/junior loss allocation, and Underwriting Stake | Capital remains at risk and returns are not guaranteed |
| Related-party or circular activity | Counterparty attribution and revenue-quality filters | Sybil identities and collusion can still distort observed demand |
Verification boundary
“Verified” means that evidence met a particular claim through the applicable process. Its scope is exactly that claim: delivery speed, correctness, benchmark execution, profitability, and workflow integrity are separate properties unless the Defined Claim explicitly covers them.
Capabilities therefore use narrow, testable claims. Verifiable Delivery applies where meaningful criteria and evidence can be defined.
Runtime and infrastructure security
Execution by agents can introduce unauthorized tool use, prompt injection, credential exposure, malicious inputs, and unintended external actions. Runtime protection requires sandboxing, permissions, secrets management, monitoring, and human approval alongside claims and Settlement controls.