35 lines (22 loc) · 1.36 KB

Non-Goals

This document lists explicit non-goals for the current pre-release security posture.

Adversarial ML guarantees

We do not claim perfect jailbreak prevention.
We do not claim complete resistance against adaptive adversarial prompt attacks.

Host compromise classes

We do not defend against kernel-level compromise.
We do not defend against malicious root/admin on the host.
We do not defend against a fully compromised dependency ecosystem.

Filesystem edge classes not universally guaranteed

We do not claim comprehensive defense across all mount/hardlink/junction corner cases on every platform.
We do not claim universal elimination of check-then-open (TOCTOU) races in all external integrations.

Platform parity limitations

Security behavior can differ by platform/runtime features and available sandboxing primitives.
Linux/macOS-oriented controls may not map 1:1 to Windows semantics.

Network perimeter non-goals

We do not claim this repository alone enforces kernel/network microsegmentation.
Defense in depth with platform/network controls remains an operator responsibility.

What we do instead

Fail closed by default for ambiguous security-relevant parsing.
Use bounded queues, inflight caps, and timeouts for DoS resistance.
Maintain targeted regression tests, fuzzing, and scheduled memory/concurrency sensors.