diff --git a/docs/security/red-team-assessment-2026-02-21.md b/docs/security/red-team-assessment-2026-02-21.md
new file mode 100644
index 0000000..5d69425
--- /dev/null
+++ b/docs/security/red-team-assessment-2026-02-21.md
@@ -0,0 +1,196 @@
+# KeyJawn Red-Team Security Assessment (Android, iOS, and whole repo)
+
+Date: 2026-02-21  
+Assessor: Codex (manual adversarial code review)
+
+## Scope
+- Android app (keyboard + SCP upload paths)
+- iOS app + iOS keyboard extension
+- Worker automation backend (`worker/`)
+- Repo-wide abuse paths affecting trust boundaries and secrets
+
+## Methodology
+- Static adversarial review of auth, secret storage, transport security, command execution, and trust boundaries.
+- Manual exploitability scoring by impact × likelihood.
+- Focused on realistic attacker models:
+  1. On-path network attacker (MitM)
+  2. Malicious local app / extension context abuse
+  3. Compromised host with config write access
+  4. Adversary with partial automation-infra access (Redis / bot integration)
+
+---
+
+## Executive summary
+The largest systemic issue is **inconsistent SSH trust guarantees**: both iOS app and iOS keyboard extension will connect with host key verification disabled when a host key is absent or invalid. That creates a practical MitM path to credential theft/session hijack.
+
+The second major issue is **private key material copied from Keychain into App Group `UserDefaults`** for extension access. This weakens key custody and enlarges secret exposure in exactly the component (`RequestsOpenAccess=true` keyboard extension) with the widest attack surface.
+
+The automation worker has a **fail-open approval mode** (auto-approve when Telegram creds are absent) and several shell-based execution chokepoints that are acceptable only if config integrity is guaranteed.
+
+---
+
+## Findings
+
+### 1) iOS SSH connections allow unverified host identity (MitM risk)
+**Severity:** High  
+**Area:** iOS app terminal + iOS keyboard uploader
+
+**Evidence**
+- App terminal falls back to `.acceptAnything()` when no pinned host key is present.  
+- Keyboard SCP uploader does the same fallback for uploads.
+
+**Attack path**
+1. Victim connects from hostile Wi‑Fi / network segment.
+2. Attacker presents rogue SSH server cert/key.
+3. Client accepts host key (because verification is disabled), attacker relays/harvests credentials and session data.
+
+**Why this matters**
+This defeats core SSH identity guarantees and enables transparent interception.
+
+**Remediation**
+- Replace fail-open behavior with strict TOFU/pinning UX:
+  - first connect: display fingerprint + explicit trust prompt
+  - later connects: hard fail on mismatch
+- Never silently downgrade to `.acceptAnything()`.
+
+---
+
+### 2) iOS private SSH key mirrored into App Group `UserDefaults`
+**Severity:** High  
+**Area:** iOS key custody
+
+**Evidence**
+- Key is stored in Keychain (good), but raw private key bytes are also written to App Group defaults for keyboard extension use.
+
+**Attack path**
+1. Attacker gains code exec in extension context / abuses open-access keyboard paths / local forensic access.
+2. Reads App Group defaults blob with private key bytes.
+3. Reuses key to authenticate as victim to configured hosts.
+
+**Why this matters**
+This bypasses Keychain protections and makes theft of long-lived SSH identity significantly easier.
+
+**Remediation**
+- Stop mirroring raw private keys to `UserDefaults`.
+- Use shared Keychain access group for extension-safe key retrieval.
+- If impossible, derive short-lived delegated credentials instead of exporting root private key material.
+
+---
+
+### 3) iOS host metadata mirrored into App Group defaults
+**Severity:** Medium  
+**Area:** iOS extension trust boundary
+
+**Evidence**
+- Host list JSON is mirrored to App Group defaults for extension reads.
+
+**Attack path**
+Tampering with host list in shared container can redirect uploads/sessions to attacker-controlled endpoints (especially dangerous when host-key verification is optional).
+
+**Remediation**
+- Integrity-protect host config (e.g., signed blob with app-held key).
+- Validate host list against strict schema and enforce pinning requirement before network use.
+
+---
+
+### 4) iOS keyboard extension requests Open Access (expected but high-risk context)
+**Severity:** Medium  
+**Area:** iOS keyboard extension exposure
+
+**Evidence**
+- `RequestsOpenAccess` is enabled.
+
+**Risk note**
+Open access is often required for network features, but it broadens data exfiltration and supply-chain impact if extension logic is compromised.
+
+**Remediation**
+- Minimize extension privileges and data residency.
+- Move sensitive operations to containing app where possible.
+- Add hard telemetry/alerts for unexpected upload/session behavior.
+
+---
+
+### 5) Android SCP client has insecure API footgun (`knownHostsManager` nullable)
+**Severity:** Medium  
+**Area:** Android transport security API design
+
+**Evidence**
+- `ScpUploader.upload(...)` accepts nullable `knownHostsManager`; when null it sets `StrictHostKeyChecking=no`.
+
+**Attack path**
+Future call sites can accidentally omit manager and silently disable host verification, reintroducing MitM risk.
+
+**Remediation**
+- Make host-key verifier mandatory in API signature.
+- Remove `StrictHostKeyChecking=no` branch.
+- Enforce compile-time secure defaults.
+
+---
+
+### 6) Worker approval path fails open when Telegram is unconfigured
+**Severity:** Medium  
+**Area:** Automation governance / posting safety
+
+**Evidence**
+- If bot token/chat id missing, `request_approval` returns `"approve"` automatically.
+
+**Attack path**
+Misconfiguration or secret-loss incident can convert gated actions into ungated auto-execution.
+
+**Remediation**
+- Fail closed (`deny`/`backlog`) when approval channel unavailable.
+- Add startup health check that refuses to run posting workflows without approval config.
+
+---
+
+### 7) Shell-based worker execution relies on config integrity
+**Severity:** Medium  
+**Area:** Worker command execution
+
+**Evidence**
+- `asyncio.create_subprocess_shell` is used in curation evaluation and social scroller command execution.
+
+**Attack path**
+If an attacker can tamper with config values (e.g., model/script path/ssh host), they may gain command injection or command substitution leverage.
+
+**Remediation**
+- Prefer `create_subprocess_exec` with argv arrays.
+- Strict allowlist validation for model names, hosts, and script paths.
+- Immutable config source + permission hardening.
+
+---
+
+## Platform posture summary
+
+### iOS
+- Strong points: main app uses Keychain for base key storage.
+- Critical weaknesses: host verification can be skipped; key material copied to weaker store; open-access extension magnifies blast radius.
+- Priority: **fix host verification and key custody first**.
+
+### Android
+- Strong points: encrypted host storage; upload path uses known-host workflow at current call site.
+- Weakness: uploader API still permits insecure branch; host-key fingerprints in regular prefs can be tampered on compromised device.
+- Priority: remove insecure branch and enforce verifier requirement.
+
+### Worker / whole system
+- Strong points: approval flow exists; HTML escaping is done for Telegram rendering.
+- Weaknesses: fail-open approvals + shell execution surfaces dependent on trusted config boundary.
+- Priority: fail-closed governance and shell hardening.
+
+---
+
+## Prioritized remediation plan
+1. **Block all unverified SSH sessions** (iOS app + keyboard, Android future-proofing).
+2. **Eliminate private key export to App Group defaults**; move to shared Keychain access group.
+3. **Fail closed on missing approval channel** in worker.
+4. Convert shell invocations to exec-style APIs and validate config inputs.
+5. Add regression tests asserting:
+   - no `.acceptAnything()` in production code paths
+   - approval request without Telegram config returns backlog/deny
+   - uploader cannot be called without host verifier.
+
+## Suggested adversarial test cases (post-fix)
+- MitM harness with rogue SSH server key to verify hard-fail behavior.
+- Extension-container tamper test for host list and key retrieval.
+- Worker boot with missing Telegram creds must halt posting pipeline.
+- Config poisoning tests for command construction paths.