Aletheia Sentinel
holeyfield33-art/aletheia-sentinel
Análise forense autônoma com cadeia de evidências criptografada
Ver no GitHubMCP server para resposta autônoma a incidentes em imagens forenses (memória/disco), com três agentes especializados (Scout, Nitpicker, Judge) que realizam triagem baseada em evidências. Cada descoberta é rastreável a execuções criptograficamente recebidas de ferramentas SANS SIFT (Volatility, Regripper, Plaso, EVTX).
Instalação
npx convez-ai holeyfield33-art/aletheia-sentinel1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Investigação autônoma de incidentes em imagens de memória
- Análise forense de discos rígidos com rastreabilidade de evidências
- Resposta a compromisso multi-host com verificação de cadeia lateral
- Auditoria forense com prova criptográfica de integridade
Veredito

Solução sofisticada e bem arquitetada para resposta forense autônoma com garantias de segurança por construção (não por prompt). Validado em datasets reais SANS com detecção correta de campanhas multi-host; ideal para equipes de resposta a incidentes e pesquisadores, mas requer expertise em forense digital.
README
A tradução para o português deste README ainda está na fila.
Aletheia Sentinel
Autonomous incident response on the SANS SIFT Workstation, with evidence-grounded self-correction -- every finding traces to a tool execution.
SANS "Find Evil!" Hackathon -- Architectural Approach #2: Custom MCP Server
What it does
Sentinel is a custom MCP server that gives an autonomous agent a typed forensic
tool surface -- volatility_pslist, volatility_netscan, volatility_cmdline,
regripper_amcache, plaso_log2timeline, evtxecmd_security -- instead of raw shell. A three-agent
loop (Scout plans, Nitpicker reviews, Judge synthesizes) triages a memory or
disk image and produces an incident report where every claim is backed by a
cryptographically-receipted tool execution.
Watch the 5-minute demo -- autonomous Scout->Nitpicker->Judge investigation on real SANS SRL-2018 evidence, with live HMAC receipt-chain verification.
Validated on real SANS SRL-2018 evidence across two hosts (rd01, wkstn-05)
linked by a shared C2 (172.16.4.10:8080) and a proven directional lateral
link (rd01 -> wkstn-05 via inbound SMB) -- a documented multi-host,
multi-technique campaign. The pipeline surfaced the real implant (p.exe in
c:\windows\temp\perfmon), the rd01 WMI->PowerShell->cmd.exe launch chain and wkstn-05 WMI->PowerShell->rundll32 beacon chain, and
caught its own false positive: a binary first flagged as a backdoor was
correctly identified as F-Response forensic tooling on rd01 from its own
network activity and command line -- a trap that recurred on both hosts,
resolved where the identifying evidence was present and held as "suspicious
pending verification" where it was absent. Both receipt chains verify. Findings were independently verified
against raw Volatility output. See the
accuracy report.
Why it's safe by construction
- Architectural guardrail, not a prompt. Destructive commands aren't in the typed tool surface, so the agent physically cannot run them. Evidence paths are pinned server-side -- the agent cannot redirect a tool at an arbitrary file, and tools are gated by which evidence was actually provided.
- Evidence-bounded classification. Identity discipline is split across two agents: the Scout must verify a flagged process's identity from evidence (command line, or correlation of its network activity to known infrastructure) before concluding, and the Judge will not report a process as malicious -- or let it drive the compromise verdict -- unless that identity is established in the findings; otherwise it is held as "suspicious pending verification." The Nitpicker is a consistency reviewer, not an evidence gate.
- Tamper-evident audit trail. Every tool execution writes an HMAC-signed, hash-linked receipt (SHA-256 of input and output). Altering any receipt breaks verification -- findings trace to executions.
- Spoliation-safe. Read-only pipeline; the source image hash is unchanged after analysis, matching the original acquisition log.
Experimental signal (advisory, off by default): an optional spectral
confidence score (via Geometric Brain MCP) can be enabled with --spectral.
A calibration study across four models found it weak and non-decisive
(AUROC ~0.5-0.7), and the live service returns INSUFFICIENT_DATA on
forensic-length samples, so it is opt-in rather than a silent external
dependency. The reliable evidence-discipline path is the Scout's
identity-verification rule combined with the Judge's evidence-bounded synthesis,
with a deterministic STRESSED->re-investigate guard retained in the
orchestrator as defense-in-depth. We report this honestly rather than claim a
detector the data does not support.
Details in the accuracy report.
Quickstart (60 seconds, no API key)
pip install -e '.[dev]'
sentinel demo # scripted investigation + verifiable receipt chain
sentinel demo runs a full scripted investigation with no API key and no SIFT
tools installed, then writes a receipt chain you can verify (the demo prints
the exact sentinel verify command and chain path).
Run on real evidence
export ALETHEIA_RECEIPT_SECRET=$(python -c "import secrets;print(secrets.token_hex(32))")
export ANTHROPIC_API_KEY=sk-ant-...
# Requires Volatility 3 (`pip install volatility3`) for memory tools.
sentinel run my-case --image /path/to/memory.img
sentinel verify audit-logs/<chain>.jsonl
Tools are gated by the evidence provided: a memory image enables the Volatility
tools; a disk image enables plaso; a registry hive enables regripper; .evtx
logs enable evtxecmd. The agent is never offered a tool whose evidence is
absent.
Architecture
graph TD
A[Scout Agent<br/>LLM: plans next tool call] -->|ScoutDecision| B[Orchestrator<br/>caps + routing + path pinning]
B -->|tool_name + args| C[MCP Server<br/>typed tool surface]
C -->|parsed ToolResult| B
B -->|append| D[Receipt Chain<br/>HMAC + hash-linked]
B -->|ToolResult| E[Nitpicker Agent<br/>LLM: consistency review]
E -->|accepted / rejected| B
B -.->|reasoning text| F[Spectral Gate<br/>advisory confidence]
F -.->|SpectralHealth| B
B -->|accepted findings| G[Judge Agent<br/>LLM: synthesize report]
G -->|SessionResult| H[Receipt-Sealed Report]
subgraph SIFT Workstation
C --> I[volatility]
C --> J[regripper]
C --> K[plaso]
C --> L[evtxecmd]
end
Solid arrows are enforced control flow; the dashed spectral path is advisory. Full trust-boundary table in docs/architecture.md; pattern overview in architecture.md.
Development
All three checks pass on this revision:
ruff check src tests # All checks passed!
mypy # Success: no issues found in 43 source files
pytest # 153 passed
Part of the Aletheia ecosystem
Sentinel applies one thesis -- enforce safety architecturally, then prove it with a tamper-evident audit trail -- to digital forensics. Sentinel runs standalone (no other Aletheia repo is required to install or run it). Related projects that share the same enforce-and-audit pattern in other domains:
- Geometric Brain MCP -- spectral analysis server (optional, advisory signal only).
- Companion projects (not required): a runtime firewall applying the pattern at module-load time, and a red-team kit as the offensive counterpart that attacks the same surfaces these tools defend.
Documentation
- Accuracy Report -- real-evidence validation, line-by-line findings verification, evidence integrity, and signing roadmap.
- Dataset Documentation -- SANS SRL-2018 provenance and per-host findings.
- Fixture Benchmark -- mocked regression benchmark with methodology disclosure.
- Architecture -- components, trust boundaries, guardrails.
License
Apache 2.0 -- see LICENSE.