Security model

Everything is
end-to-end encrypted.

MirrorClaw is a secure agent workspace built around three ideas: keep the primary inference path on the Mirror-protected route, move longer or riskier work into an isolated runtime boundary, and make proof visible inside the product.

E2E
Encrypted route
AES-256-GCM via per-request CEK
205
Lifecycle cases
MirrorArena security bench
75
Workspace scenarios
ClawSafety-style harm bench
15
Trust-gap cases
12 trust-boundary families

Trust boundaries, all operator-visible

MirrorClaw separates the product into distinct trust boundaries. Each one has a surface in the workspace.

Protected access

Token-gated /app and onboarding gate

Hosted alpha access is explicitly gated. The workspace doesn't default to a public unauthenticated chat surface.

Mirror inference route

Request, route, response, and decrypt receipts

Encrypted prompt and response transport on the main model path. Live evidence exposed in the product.

Workspace thread

Chat lane, proof rail, Show proof affordances

Conversation, memory, and proof in one operator surface. Proof attached to chat, not buried in backend logs.

Delegated runtime

Jobs, Stack, and delegated task receipts

Long-running or tool-heavy work moves off the main thread. Reviewable through Jobs and Stack.

Connector boundary

Readiness, extensions, and setup posture

Custody and readiness made visible instead of implying trust. Hosted blockers and dependency gaps surfaced.

Release gate

MirrorArena release lanes

Hosted builds are pressure-tested through MirrorArena before they ship. Security evaluation as part of shipping.

The receipt chain on every turn

What an operator actually sees in the workspace, attached to the chat surface.

Enc in

Prompt is wrapped with a per-request CEK before transport. The gateway sees ciphertext.

Route

Request forwarded over the Mirror inference path. Route, policy, and runtime lane logged.

Enc out

Model response returns on the encrypted lane. Same CEK reference, same boundary.

Decrypt

Decryption happens inside the trusted workspace. Operators can inspect the full chain.

What MirrorArena covers

MirrorArena is the security validation harness in the Mirror stack. Seven scan families in the release loop before a hosted build ships.

205 cases

Lifecycle bench

End-to-end runtime misuse across input, planning, tool use, persistence, and exfiltration.

75 scenarios

Workspace bench

ClawSafety-style domain harm: destructive actions, secret leakage, unsafe destination changes.

15 cases · 12 families

Trust-gap bench

Ingress auth, replay, compaction poisoning, subagent trust, memory drift, identity bleed, tool-risk guards, permission integrity, and context-judge decisions.

Native red-team

SkillAttack

Skill injection, prompt seeding, exploit refinement, seeded-vs-baseline attack deltas.

Translated corpora

PromptSecurity

Jailbreak-style attacks, transformed prompt corpora, operator-scenario prompt-defense evaluation.

Repo + workflow

Agent scan

Static and agentic scanning of repo trust boundaries, provider config, MCP exposure, skill provenance.

Live target

Infra scan

HTTP target fingerprinting, version matching, advisory-driven exposure analysis.

Defense in depth

Four layers. Each one independently testable, each one logged.

Encryption

  • AES-256-GCM via per-request CEK
  • Request encrypted, response encrypted
  • OS keychain-rooted master key
  • Mirror gateway sees ciphertext only

Isolation

  • OpenShell short-lived containers
  • Landlock filesystem restrictions
  • Network namespace isolation
  • WASM capability-based tools

Egress control

  • Per-tool HTTP allowlist
  • TLS-inspecting proxy
  • DNS pinning for SSRF protection
  • Approval gate for risky calls

Audit

  • Per-turn trace (route + policy + tools)
  • CEK references in evidence bundle
  • Trace, policy, ATIF exports
  • Release-gated by MirrorArena

Three deployment shapes

Same product story, different boundary separation.

Hosted alpha
Early design partners and guided onboarding

Protected route, visible proof, shared hosted control plane

Workspace tier
Small teams needing a productized secure agent workspace

Encrypted inference included, readiness and connector policy visible

Dedicated tier
Customers that need stronger isolation and deployment control

Dedicated VM or confidential-runtime path with stronger boundary separation

Security as a visible product surface.

Other agent workspaces ask you to trust their backend. MirrorClaw shows you the receipt.