Security Overview
Runtime Governance and Environmental Isolation
1. The Core Principle: Capability != Authority
In conventional agent frameworks, discovering a tool automatically gives the model unchecked power to invoke it. Apoc breaks this vulnerability by enforcing an authoritative security boundary outside the reasoning model. An agent may know that a deletion, deployment, or shell tool exists, but the runtime kernel prevents execution without explicit policy validation or interactive human authorization.
2. Hardware-Enforced Sandboxing
Workflows run in segregated subprocess workers with bounded filesystem and network paths. High-impact commands cannot escape the designated workspace boundary, mitigating malicious prompt injection attacks that attempt to modify root files or leak unauthorized documents.
3. Cryptographic Human Sign-Off
Operations classified as high-risk (such as production deployments, irreversible file rewrites, or credential access) trigger a physical permission gate. The system presents the exact target, arguments, and risk classification to the user before any action is committed.
4. Evidence-Driven Outcome Verification
Apoc never relies on conversational claims or model confidence scores to confirm success. Every action must produce measurable physical evidence: verifying that sockets respond, files exist with valid checksums, and test suites exit with zero errors.