I Tried to Sneak Four Bad Agents Past My Own Certification Gate. All Four Got Blocked.

Описание и применение

Автор протестировал систему HivePlane, создав четыре агента, пытавшихся обойти контроль сертификации. Все атаки были успешно заблокированы, подтверждая эффективность проверки сертификатов, привязки модели, поведения агента и лимитов затрат.

Задача

Ensuring that only certified agents can run in production and that various attack vectors (uncertified registration, model swap, regression, and cost overrun) are blocked.

Результат

The certification gate successfully blocked all four attack scenarios, validating admission control, model binding, behavior checks, and cost limits.

Рабочий процесс

• Register agent without certification and attempt production run – gate blocks with 403.
• Certify agent on one model, run on a different model – gate blocks with 403.
• Run a regression agent that meets output format but fails behavior checks – gate blocks with 403.
• Run an agent that exceeds per-run cost ceiling – gate blocks with run budget exceeded error.

Доказательства

run for workload 'uncertified-agent' refused admission to production: certification status 'uncertified' is insufficient for production; requires 'certified'

Источник

The run's identity is compared against the attestation's bound model, not the manifest's declared one.

Источник

Certification ran it against the production threshold and returned 201 with status 'uncertified' — correctly refusing to certify

Источник

The run failed with run budget exceeded the moment the usage report crossed the line. Recorded cost: $2.85.

Источник