This Trust Center maps the xYz agent-security architecture, threat model, public controls, validation milestones and disclosure guidance. It is designed for business owners, engineers, security reviewers and other agents that need more than a marketing sentence.
Este Trust Center mapeia a arquitetura de segurança dos agentes xYz, modelo de ameaças, controles públicos, marcos de validação e orientação de divulgação. Foi feito para empresários, engenheiros, revisores de segurança e outros agentes que precisam de mais do que uma frase de marketing.
The xYz model separates who an agent is from what it is allowed to do. A consequential request moves through explicit trust and authority boundaries before execution.
O modelo xYz separa quem é o agente daquilo que ele tem permissão para fazer. Uma requisição relevante atravessa fronteiras explícitas de confiança e autoridade antes da execução.
The matrix below does not claim that risk disappears. It shows the control boundary xYz expects to interrupt each class of agentic failure and where the public evidence can be inspected.
A matriz abaixo não afirma que o risco desaparece. Ela mostra qual fronteira de controle a xYz espera que interrompa cada classe de falha agentic e onde a evidência pública pode ser inspecionada.
| Attack | Primary control | Expected boundary | Evidence |
|---|---|---|---|
| Prompt Injection | Authority + data boundary | Prompt content cannot grant new authority. | Lab · Authority |
| Privilege Escalation | ask_first | Spending and consequential actions require approval. | Authority |
| Tool Abuse | Tool authority boundary | Tool availability does not imply permission to mutate Production. | Lab |
| Approval Bypass | Explicit approval gate | Approval-required actions remain blocked without approval. | Lab |
| Data Exfiltration | never_autonomous boundary | Private customer data is not autonomously releasable. | Authority |
| Agent Impersonation | P-256 signed identity | Unsigned identity does not satisfy trusted-agent authentication. | Protocol |
| Replay Attack | Timestamp + nonce + replay window | Old signed requests cannot be replayed indefinitely. | Protocol |
| Excessive Agency | ask_first + never_autonomous | High-impact agency is independently bounded. | Lab · Authority |
Human-readable surfaces sit on top of machine-readable endpoints. The raw protocol remains available for technical review and automated inspection.
As superfícies legíveis ficam sobre endpoints legíveis por máquina. O protocolo bruto continua disponível para revisão técnica e inspeção automatizada.
This log records xYz-owned validation milestones. It does not publish or identify visitor challenge activity from the Security Lab.
Este histórico registra marcos de validação executados pela própria xYz. Ele não publica nem identifica a atividade de visitantes no Security Lab.
Approval Bypass rendered BLOCKED with the ask_first approval gate and live Authority Map evidence.
Added safe challenge simulations for prompt injection, privilege escalation, tool abuse, approval bypass, data exfiltration, impersonation, replay and excessive agency. Production rendering was independently revalidated after a JavaScript regression fix.
Added a human-readable live view over identity, Authority Map, network trust, signing self-test and consequential-action controls while preserving raw JSON evidence.
Added a public deterministic assessment surface for agent identity, authority, approvals, tool access, secrets handling, replay protection, kill switch, isolation and data boundaries.
Security claims remain tied to public evidence. A change to a control should be reflected in the Inspector, Trust Center, tests or machine-readable trust manifest instead of living only in marketing copy.
For a non-sensitive first contact, use the xYz website conversation entry point and clearly identify the message as a security report. Do not include exploit secrets or sensitive material in the first message.