Sentinel 3.4.0

Sentinel 3.4.0

Changed

• Define standalone userland protection as the product target, with Defender optional. Publish implemented mechanisms, missing EPP capabilities and acceptance priorities in the capability matrix; do not claim full antivirus or every-product parity.
• Separate VBS, Credential Guard and Memory Integrity runtime service states from configuration, and query verified-system-image LSASS process protection with query-only access. Audit vulnerable-driver blocklist configuration without claiming enforcement. Explain Credential Guard's domain-credential scope; inactive CG is contextual rather than a universal personal-PC failure. Correct the old configuration-only IOMMU claim and add explicit inactive/unsupported dashboard labels.
• Bind persistent TCP/drop/DNS history to process creation time, reject unavailable/reused identities, and report TCP collector failures through sensor health. DNS retry bursts remain observations rather than standalone containment authority.
• Restrict HKLM Run/RunOnce cleanup to exact direct launches of verified quarantined files; recheck the typed value before deletion and verify absence afterward. Preserve argument-only references, changed values and unverified ancestors. Log cleanup attempts; cross-process registry mutation is not atomic.
• Match injected-victim composite requests to actual scoped process containment so verified exits do not falsely require backing-file quarantine. Add native Windows fixtures for real owned-process exit, unrelated-process survival, temporary registry attribution and runtime-provider reporting.
• Fix the credential-read regression fixture that assumed a fabricated PID had a live Windows identity. Inject controlled birth times into the production rule/correlator and verify positive corroboration, missing identity and PID reuse without weakening production checks.
• Extend injection-candidate PE inspection to arbitrary host identities and RX code with readonly allocation headers. Validate executable PE sections with bounded read-only walks; MZ/JIT/ordinary image mappings do not provide payload proof. Browser/API suppression still queues PE-only target inspection without granting response authority.
• Carry verified victim creation time through the PE-plus-credential-read chain to a scoped containment adapter. Reject PID reuse; preserve the victim's backing EXE and parent, skip ISO/container/VM mutation and prevent reinfection-cache poisoning from an injected clean host. Independent loader/module responses remain available. See DLL-loader and payload scope.
• Extend the existing raw-socket response rule with a bounded inline-Python signature: a connected socket supplies all three standard streams of an explicitly invoked shell. Python, Ren'Py, installer names and relay labels alone provide no response authority.
• Reuse the PowerShell decoded-code parser in the raw-socket rule, closing a path that could mistake ordinary TCP clients or quoted payload documentation for a reverse shell.
• Correlate a real Kernel-File credential-store read with confirmed C2 or reverse-shell evidence in the same live process lifetime. Require fresh timestamps and a matching creation time; watcher notifications, unknown/reused identities, weak network observations and consultant signals cannot complete this chain. Prevent the broader session-theft rule from bypassing these checks for raw store-access observations.
• Include ready removable volumes alongside fixed secondary drives in file watching and initial scanning. Reconcile volume attachment/removal within the scanner, releasing old watchers and restarting the lazy walk without consuming monitor failure-restart attempts. Existing file exclusions, scan budgets and observe-only file-verdict policy remain in force.

Scope and validation

• Keep NanoAV source and its response rules unchanged. No Ren'Py-specific blocklist or unidentified “Ome-Relay” signature is introduced; neither forum post supplied a sample or detection name. See installer and relay coverage.
• Release product projects and shipping tests cross-build with 0 warnings and 0 errors. The optional Windows Desktop ML trainer was not validated on Linux. The expanded selected portable regression suite passed 255 cases; five Windows-only fixtures were skipped. All 26 dashboard checks pass. The Windows ACL-dependent response test and native Framework runner remain unrun here. NanoAV supplied-source verification passes.
• Tests use inert telemetry and check reverse-shell dispatch authorization, credential-chain corroboration, ordinary networking, legitimate game launches, unrelated streams, stale/reused identities and local-volume selection. Add native readonly-header/RX memory and clean Qt policy fixtures, plus fake-handle victim containment/parent-preservation tests. A dedicated Windows CI job now requires all five endpoint fixtures to execute and pass without skips. Native Windows installer, ETW attribution, removable-drive lifecycle and actual containment checks remain required; portable tests do not establish coverage against either forum sample.