convez.ai
MCP ServerSegurançaAvançado
aryarahimi1

SEALED - Servidor MCP Forense com Custódia Criptográfica

aryarahimi1/sealed

Investigação forense autônoma com cadeia de custódia criptograficamente selada

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP especializado em análise forense (DFIR) que integra ferramentas reais como Volatility 3, plaso e Sleuth Kit com garantia de cadeia de custódia assinada em ed25519. Cada achado passa por um portão de proveniência que valida contra dados realmente coletados, rejeitando alucinações e garantindo admissibilidade em tribunal. Ideal para investigações onde a evidência digital precisa ser incontestável.

Instalação

$npx convez-ai aryarahimi1/sealed

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Triagem e análise forense de hosts comprometidos com prova admissível em tribunal
  • Construção de timeline unificada entre múltiplas fontes de evidência forense
  • Investigações de incidentes onde a integridade e rastreabilidade da cadeia de custódia são críticas

Veredito

Ferramenta altamente especializada e bem arquitetada para DFIR defensivo em escala. A abordagem de cadeia de custódia assinada e proveniência verificável é rara e valiosa; requer profunda expertise em segurança ofensiva, forense digital e verificação criptográfica para extrair valor completo.

README

A tradução para o português deste README ainda está na fila.

SEALED

An autonomous DFIR agent whose every finding is admissible evidence, not a SOC ticket.

Attackers now break out in as little as 27 seconds (CrowdStrike 2026 Global Threat Report, down from 51s a year earlier), and the offense is already an AI: in Nov 2025 Anthropic reported GTG-1002, a state-sponsored group that drove an AI coding agent through a campaign against ~30 organizations, with the AI executing 80–90% of the tactical work autonomously. Defense has to answer at machine speed, but an agent that hallucinates produces evidence no court will accept, and evidence excluded for a broken chain of custody re-runs the whole investigation.

SEALED is a purpose-built typed, read-only Model Context Protocol server over real forensic tooling (Volatility 3, EZ Tools, plaso, Sleuth Kit), driven by an autonomous agent. It triages a host, builds a cross-source timeline, and seals findings, but it cannot call a finding CONFIRMED in prose. Every finding must pass a provenance gate, and the whole run lands as a cryptographically signed chain of custody. Read-only mounting was table stakes in 2008; what decides cases is chain of custody, Daubert admissibility, and a documented error rate, and that is the layer SEALED is built around.

What makes it different

  1. No code path can mutate evidence. The typed surface exposes only read/parse functions (registry_query(), evtx_logons(), vol3_pslist(), mft_timeline(), …). There is no execute_shell_cmd, no write/delete tool, no shell escape. A destructive command, even one injected into a log line in the evidence, has nowhere to run, because the capability does not exist as a code path.
  2. Signed, tamper-evident custody ledger. Every tool call SHA-256s its input evidence and its output and prev-hash-chains them into an append-only ledger. Any post-hoc edit or reorder breaks the chain; trailing truncation is caught by an out-of-band length+head anchor; and the head is ed25519-signed. Verification needs only the operator's public key, published out-of-band, so an attacker who rewrites or shortens the ledger must also forge a signature under that key. An independent verifier (python -m sealed_mcp.verify) re-derives the chain, the anchor, and the signature, so you never have to trust the system that wrote the log.
  3. A provenance gate. A finding seals CONFIRMED only if the ledger verifies and every relied-on field appears in the rows the tool actually recorded, checked against the hashed bytes, never against a value the agent supplied. Otherwise it downgrades to INFERRED with the exact missing provenance.

Results on the NIST Hacking Case

On the NIST CFReDS Hacking Case (the Greg Schardt / "Mr. Evil" wardriving case), SEALED produces 11 of 12 court-sealed findings, 0 hallucinated, scored against the NIST published answer key. The 12th is deliberately abstained, because the tool emits per-account rows, not a total it can prove: the architecture would rather say "I can't seal that" than guess. The accuracy number is machine-verified, the benchmark re-runs each CONFIRMED finding back through the real gate over the rows the tools recorded, and a test asserts the committed result is reproducible from the build.

Honest scope

The gate guarantees provenance-completeness, not correctness. A model can still misinterpret real tool output, and that claim would carry a complete chain. So the claim made everywhere is:

"No CONFIRMED finding can exist without a verified tool-trace. The gate blocks unsupported assertions, not all incorrect ones."

Both directions are measured in the accuracy report: unsupported-CONFIRMED (0, re-verified by re-adjudication) and the over-refusal rate. Claiming "hallucination is architecturally impossible" would be false, so SEALED does not claim it.

Read-only is proven, not asserted

Typed tools shell to real binaries that open the image, so integrity is enforced by hashing the evidence before and after every call and recording both in the ledger. Integrity is identical before/after SHA-256, not merely "there is no shell function."

Run it

No SIFT install is needed: simulate mode runs the full pipeline over NIST-derived fixtures.

pip install -e .
python3 -m sealed_mcp.triage          # ONE command -> the full case bundle in ./sealed_out/
python3 -m sealed_mcp.redteam         # 9 attacks on the architecture, all blocked, exit 0
python3 accuracy/run_benchmark.py     # accuracy re-verified by replaying findings through the gate
python3 -m sealed_mcp.verify \
  sealed_out/custody-ledger.jsonl sealed_out/custody-attestation.json   # verify custody yourself
python3 -m sealed_mcp.smoke           # read-only tool surface + structured results

sealed_mcp.triage writes an interactive, self-contained dashboard.html (timeline, sealed findings beside the NIST answer key, the signed custody chain, the red-team panel, and the error rates; opens offline with a double-click, no server or build step), a court-ready case-report.html of the same case, a cross-source timeline.json, findings.json, Daubert provenance cards/, the hash-chained + ed25519-signed custody-ledger.jsonl with custody-attestation.json, and a gate-verified accuracy-report.json. sealed_mcp.redteam injects a destructive command, forges seals three ways, tampers a signed ledger, and tries to overwrite source/evidence; the architecture refuses every one.

The MCP surface

The agent can only ever call what is registered: typed read-only forensic tools plus a small set of integrity tools. Nothing destructive exists to call.

ToolPurpose
registry_query, sam_accounts, parse_ini_artifact, image_hashRegistry hives, SAM accounts, text/INI artifacts, acquisition-hash verification
evtx_logons, vol3_pslist, vol3_netscan, mft_timeline, amcache_programs, prefetch_runsLogons, processes, network, $MFT, installed programs, execution
timelineCross-source, chronologically-merged, provenance-cited event stream
seal_findingAdjudicate a claim -> CONFIRMED or INFERRED + provenance card
verify_custody_ledger, attest_custody_ledger, get_custody_entriesVerify the chain, sign the head, introspect the ledger
provenance_pdf, export_custody_ledger, triage_statusDaubert card export, execution-log export, scope/injection-surface check

The server also exposes MCP resources (sealed://doctrine, sealed://tool-catalogue, the live sealed://custody-ledger, sealed://ground-truth/nist, sealed://accuracy-report) and prompts (/triage_host, /forgery_demo). Run /forgery_demo to watch the gate refuse a fabricated CONFIRMED finding live, in your own environment. See docs/MOMENT.md.

Documentation

  • ARCHITECTURE.md — the trust boundary and where each guarantee is enforced
  • DEPLOY.md — local run and the SIFT Workstation OVA setup (SEALED_SIMULATE=0)
  • accuracy/ — the gate-verified benchmark against the NIST answer key
  • docs/MOMENT.md — the forgery-rejection demo
  • evidence_docs/ — dataset documentation

License

MIT, see LICENSE. Built as an upstream-able contribution to Protocol SIFT, and developed for the SANS Find Evil! challenge.

Ver o original no GitHub

Bom para

Relacionados