Autonomous Digital Forensics & Incident Response agent built on Protocol SIFT. Submission for the FIND EVIL! hackathon (SANS/GIAC, Apr 15 – Jun 15, 2026).

APTWatcher is a defensive AI agent that triages compromised hosts at machine speed, correlates forensic artifacts across multiple intelligence sources, and produces analyst-grade, audit-backed IR reports — without hallucinating and without ever touching the evidence.

In 2025, the GTG-1002 campaign demonstrated a state-sponsored operation in which an agentic LLM executed roughly 80–90% of the attack chain — recon, initial access, lateral movement, data staging — without human keystrokes. The traditional defender loop (a ticket, a human analyst, a multi-day triage window) cannot match an adversary that completes its objective in minutes. See the threat model in docs/SCOPE.md.

APTWatcher's answer is parity of speed, not parity of autonomy. It runs the same primitives the attacker does — agentic orchestration, an MCP tool surface, iterative plan-execute-verify loops — but constrained by a strict read-only evidence mode, consent-gated state changes, and an append-only signed audit log. The defender keeps authority over chain of custody and publication; the agent supplies the speed.

One idempotent installer for any Linux host with Python 3.11+ (a SIFT workstation gets the full toolchain probed automatically). It refuses to run as root and never writes outside ~/APTWatcher:

curl -fsSL https://raw.githubusercontent.com/aptwatcher/APTWatcher/main/install.sh | bash
source ~/APTWatcher/.venv/bin/activate

Three commands that work immediately — no forensic VM, no API keys, no real evidence:

aptwatcher version    # build and environment info
aptwatcher profiles   # list the bundled triage profiles
aptwatcher eval       # run the accuracy harness against shipped fixtures

The full ten-minute judge-facing walkthrough — dry-run triage, analysis bundle, stub publication, self-scoring — is in docs/TRY-IT-OUT.md.

APTWatcher ships as three deployment modes sharing one brain:

• Mode A — Direct Agent Extension for Claude Code / OpenClaw users
• Mode B — Custom MCP Server for strict architectural guardrails
• Mode C — Hybrid combining both (recommended)

The same curated knowledge base, intel adapters, audit logger, and self-correction prompts back all three modes. See docs/design/deployment-modes.md.

Capabilities are tiered and opt-in:

Tier 0 is read-only by construction; tier 1 and above require an explicit consent_granted audit event before any tool fires. The full design is in docs/ARCHITECTURE.md, and the operating contract the agent itself runs under — the orchestrator brain, with the Prime Directive, role allow-lists, and failure-mode playbook — is CLAUDE.md.

• demo/SCRIPT.md — the live demo script.
• docs/demo/WALKTHROUGH.md — annotated end-to-end walkthrough with expected output.
• docs/scenarios/ — rubric scenarios S01–S03 (single Windows compromise, multi-host lateral movement, ransomware pre-detonation).
• scenarios/ — demo scenarios S04–S05 (signed offline-to-online bundle handoff; one-shot E01 triage to a signed IncidentBundle on a 14.5-minute budget).

For hackathon judges, the four submission documents in one place:

• docs/TRY-IT-OUT.md — step-by-step local run against the shipped fixtures (no live deployment by design: read-only, offline).
• docs/DATASET.md — evidence dataset inventory, provenance, and clean-room policy.
• docs/ACCURACY.md — accuracy methodology, baseline, and honest limitations.
• docs/DEVPOST.md — the project narrative.

• Read-only by default. All 42 tier-0 forensic tools never modify evidence and never emit outbound traffic. Derivatives go to work/ and out/, never next to the artifact.
• Consent gating. Tier 1+ tools require a matching consent_granted audit event, enforced in the MCP server — a tool cannot bypass it. See docs/design/tier-gating.md.
• Dry-run publication. All five adapters default to dry_run=True; going live requires the --live flag plus a signed IncidentBundle.
• Signed audit log. Every plan/execute/verify/self-correct step emits an append-only, signed AuditEvent with token telemetry; render it with aptwatcher audit-render. Findings cite event IDs or they are not findings.
• Signed bundles. Ed25519-signed IncidentBundles carry results across the offline-to-online gap, with tamper detection on import. See docs/design/evidence-integrity.md.
• No shell escapes. Tools invoke binaries by absolute path with arg-vector calls — no shell=True, no eval, no string interpolation into a shell.

The DFIR qualifier exists to disambiguate this project from a similarly named one. APTWatcher DFIR (this project) is a defensive digital-forensics and incident-response agent. APTWatch (aptwatch.org) is a separate threat intelligence platform that exposes api.aptwatch.org. APTWatcher consumes that API as one of its MCP tools — they are independent projects under the same brand family. The package, CLI (aptwatcher), and repository keep the short name APTWatcher; APTWatcher DFIR is the full display name.

In development. See SUBMISSION-CHECKLIST.md for the pre-submission gate and the GitHub issue tracker for current work.

MIT — see LICENSE.