Memory Firewall

mode …

Prompt injection against a SOC triage agent, run twice side by side. The left column has no defenses; the right has the ones you pick.

Defenses for the right column: D1 untrusted tagging D2 memory gate D3 skill policy

The eval in replays/eval.md measures D1 alone, because D2 and D3 are deterministic code already proven in tests — only D1 depends on how the model behaves.

Undefended

Pick a scenario and press Run.

Defended

Same attack, defenses on.

Quarantine

Facts held for human review — nothing here is recalled by a later run.

Nothing saved yet.