Autonomous DFIR App — Standalone · Plugs into AEGIS-AI

Every claim
re-executed. Every byte cited.

SLEUTH runs a complete digital-forensics investigation autonomously — raw disk, memory, and network evidence to a structured, fully-cited report, with no human in the loop. Run it as a standalone app, or plug it into AEGIS-AI as the SOC's tamper-evident evidence broker. The guardrail is an architecture, not an instruction.

⚠ INVESTIGATION STREAM SIMULATED · ALL FIGURES MEASURED — SEE docs/ACCURACY.md
79
Confirmed Findings
LONE-WOLF CASE · 0 REFUTED
161
Validator Re-Executions
EVERY CLAIM RE-RUN
<100ms
Full Provenance Cite
es cite F-NNN
10/18
Self-Corrections Recovered
HYPOTHESIS → RETRY → VERIFY

Findings Stream — lone-wolf caseINVESTIGATING

Lateral-Movement Attack GraphAGE

VPN entry wkstn-42 T1059 svc-acct T1071 fileserver T1048 backup
APACHE AGE NODES · EVERY EDGE CARRIES A FINDING CITATION

Merkle-Chained Tool CallsVERIFYING

Agent ActivityRUNNING

The Guardrail Is the Architecture

Bash PreToolUse hook — exits 1 on anything that isn't ./bin/sb or ./bin/es  strings → DENIED · grep → DENIED · tshark → DENIED
Rust broker mediates every call — pg_jsonschema argument validation, stdout/stderr streamed to Postgres
Rootless podman + custom seccomp — each forensic binary isolated, read-only evidence mount
Validator re-execution — every finding's original tool call re-run with identical args; unreproducible → refuted, excluded from report
Postgres is the brain — every tool call is a row, every artifact content-addressed (BLAKE3), pgvector cosine dedup on IOCs

Standalone ModeAPP

One command, full investigation — ./scripts/investigate.sh ./cases/lone-wolf/: triage → 3 specialists → validator → cited report.md
Browser inspector included — compiled Rust SaaS on :8932: cases, findings, audit chain, attack graph. No notebook, no terminal.
Bring your own evidence — SANS M57-Patents / LoneWolf today; any disk image, memory dump, or pcap tomorrow.

Plugin Mode — inside AEGIS-AIsoc-dfir

EvidenceBroker interface — soc-dfir wraps SLEUTH as the sandboxed broker: the only thing allowed to touch the isolated capture store.
Findings become SOC evidence — every DfirFinding carries a BLAKE3 Citation; verify_citation recomputes it against the live store — spoofed or stale findings are rejected, not trusted.
Agents that consume it — agent-investigate escalates cases to SLEUTH; agent-hunt sweeps its indicators; citations ride the same keyed-BLAKE3 audit chain as every SOC action.

Hallucinations Caught — Self-Correction Log

SpecialistFailedCorrectionResult
memorylinux.pslistderive OS profile → windows.info, re-run Windows pluginscaught + fixed
networktshark (truncated PCAP −4096B)editcap_recover → recovered.pcap → re-run tshark6/6 runs
memoryprofile hint spoofed (corruption test)validator re-executed linux.pslist, confirmed still failsrefuted = 0 trusted
10/18 SELF-CORRECTIONS SUCCEEDED — FAILED HYPOTHESES WERE RETRIED, NEVER TRUSTED UNVERIFIED.

Persistent Inspector (SaaS) — :8932

/
Case list — evidence counts, last-activity badge
/case/<id>/findings
Every finding — confidence, validation status, MITRE tag
/finding/<fid>
Tool call, exact args, stdout hash, artifact, Merkle root
/case/<id>/audit
Merkle-chained calls with live verify badge
/case/<id>/graph
Apache AGE attack graph — lateral movement at a glance
es cite F-001
Full provenance chain in <100ms — judges run live SELECTs

Three specialists on the full SIFT toolchain

Disk, memory, and network specialists run in parallel over real SANS evidence (M57-Patents / LoneWolf 2018), then a validator re-executes every claim and an IR-narrator writes the report — every sentence carrying a [F-NNN] citation back to provenance.

💾

disk

Sleuth Kit + Plaso — filesystem timeline, MFT, tsk_recover. 34 confirmed findings on the primary case.

🧠

memory

Volatility 3 — process injection, SSDT hooks, OS-profile derivation with hallucination recovery. 21 confirmed.

🌐

network

tshark + Zeek + Suricata (ET-Open) — beacon detection, truncated-PCAP recovery. 24 confirmed.

🔗

Tamper-Evident by Construction

BLAKE3 content-addressed artifacts, Merkle-chained tool_calls hypertable, pgaudit on every SELECT. es cite F-NNN returns the whole trace.

⚡

Event-Driven Re-Validation

Per-finding DB trigger + LISTEN/NOTIFY — no clock, no human input. Self-correction bounded to 3 retries per finding.

🧔

Rust + Postgres 17

Memory-safe broker, Apache AGE attack graph, pgvector RAG grounding, pg_jsonschema argument validation.

Pipeline

SANS evidence (disk · memory · network) │ ┌──────▼──────┐ │ TRIAGE │ classify evidence, build case substrate └──────┬──────┘ ┌───────────┼───────────┐ ┌───▼───┐ ┌───▼────┐ ┌───▼────┐ │ DISK │ │MEMORY │ │NETWORK │ Sleuth Kit + Plaso · Volatility 3 · tshark/Zeek/Suricata └───┬───┘ └───┬────┘ └───┬────┘ └───────────┼───────────┘ ┌──────▼──────┐ every claim re-executed with identical args │ VALIDATOR │ confirmed / refuted / inconclusive / drift └──────┬──────┘ unreproducible → excluded from report ┌──────▼──────┐ │ NARRATOR │ report.md — every sentence carries [F-NNN] citation └─────────────┘ best-of-N · Opus judge PLUGIN MODE — inside the AEGIS-AI substrate: agent-investigate / agent-hunt → soc-dfir::EvidenceBroker → SLEUTH broker (isolated capture store, rootless podman) → DfirFinding + BLAKE3 Citation → verify_citation() → verified evidence joins the SOC's hash-chained audit trail
Built on
SANS SIFTSLEUTH KITPLASOVOLATILITY 3TSHARKZEEKSURICATA ET-OPENPOSTGRES 17APACHE AGEPGVECTORBLAKE3ROOTLESS PODMANAPACHE-2.0
Read the Code Talk to Us See AEGIS-AI