Dagda is an in-development command platform built around that question: one AI layer designed to work across hundreds of security tools and 9 operational modes — monitoring, investigating, setting traps, and responding — inside a safeguard system it isn't able to loosen on its own.
Security work often spans dozens of separate tools. Different CLIs, different output formats, no shared context. Your nmap scan doesn't know about your Suricata alerts. Your forensics don't inform your hardening. Every context switch is a gap an attacker can exploit.
Most alerts are noise. Without cross-source correlation, every anomaly looks the same. You're context-switching between dashboards, losing the thread of actual attacks while chasing false positives. The real breach hides in the flood.
Senior analysts are expensive and scarce. Junior analysts need months of mentoring. Offensive skills atrophy without practice environments. Knowledge walks out the door with every resignation. Your team's ceiling is your most experienced member.
"The question behind Dagda: what if the tools you already have could share context?"
Named for the Irish god who carried a club that could kill with one end and heal with the other. The aim is a platform that does more than detect: it's designed to correlate, respond, and teach — and to explain its reasoning as it goes.
Claude API primary, Ollama offline fallback. Designed as an agent rather than a chatbot: hundreds of tools, a knowledge graph, Bayesian threat networks, and temporal correlation, with distributed Crow agents to extend it across a fleet. The goal is a system that can respond, not just alert — and the hard part is making that trustworthy.
Bug bounty hunting, purple team training, honeypot operations, and automated incident response — all behind a safeguard system designed to make misuse structurally hard, not just against the rules. One platform for red, blue, and purple work.
The Ollama local-LLM fallback is meant to let Dagda run with no cloud connection at all, in air-gapped labs and isolated networks. No cloud dependency and no telemetry by design: your data stays on your machine.
Educational layer with 8 attack anatomies, 5 guided investigation workflows, and an Arena mode that turns real detected attacks into training scenarios — so practice comes from your own environment, without leaving the platform.
Each mode is designed to unlock a different capability set with its own safeguard requirements; eight are previewed below. Switching modes changes which tools the AI suggests, how much risk it will tolerate, and how it carries itself.
Monitors CPU, memory, disk, network, and processes in real time. Detects brute force attacks, ARP spoofing, rogue DHCP, port scans, and DNS anomalies. Responds autonomously — blocks IPs, kills suspicious processes, quarantines files — then explains exactly why it acted.
SOC Level 1 workflow. Assess alerts quickly, determine severity, escalate or close. Speed and accuracy matter most — don't investigate deeply here. Escalate to INVESTIGATOR for analysis or INCIDENT COMMANDER for response.
OSINT, disk and memory forensics, evidence chain of custody, threat intel correlation. Every artifact is hash-verified, every finding cross-referenced through the Knowledge Graph. Passive by default — observe without disturbing.
Deploy honeypots, canary tokens, fake credentials, breadcrumbs, and decoy files. Track and profile attackers in real time. Extract TTPs, credentials, and payloads. Disruption actions (sinkhole, tarpit) require the counter-intel toggle.
Every packet checked against your target whitelist. Nuclei scanning, SQLi/XSS/LFI/SSTI testing, API security (GraphQL, JWT, OAuth), and web crawling — all with responsible disclosure workflows built in. DNS resolution verified to prevent scope escape.
Full offensive capabilities in network-isolated VMs. 5 pre-built multi-VM scenarios. PvP (red vs blue) and PvT (player vs target) game modes with scoring, leaderboards, and AI coaching. Real attacks detected by Sentinel auto-generate new training scenarios.
Full incident response lifecycle. Case management, evidence collection, containment actions, timeline reconstruction, and response playbook execution. Coordinate from detection through remediation.
Built to keep a low profile. Anti-detection layer with timing jitter, SOCKS5 routing, user-agent rotation, and DoH rotation. Cryptographic evidence vault with SHA-256 Merkle trees and Bitcoin timestamp anchoring. Secure session cleanup with multi-pass overwrite.
Five safeguard levels are designed to control what the AI can do. Raising safety is meant to be instant and unrestricted; lowering it requires deliberate, auditable escalation. The hardest design problem is the AI itself: it shouldn't be able to loosen its own constraints.
More than a wrapper around an LLM: the design pairs a Rust-enforced permission boundary with intelligence-fusion engines and a broad set of integrations.
Permission gate, sandbox engine, audit logger, and crypto module — all in Rust via PyO3. Memory-safe, zero-cost abstractions, tamper-resistant. Python fallbacks for cross-platform development.
A cognitive architecture rather than a chatbot wrapper: persistent memory, probabilistic reasoning, and cross-module intelligence fusion across 13 bridges.
Designed to work with the tools security teams already use. Each integration follows the same pattern: detect, health-check, run, parse, ingest.
Dagda isn't publicly released yet. This is the setup flow it's being built around.
9-step guided configuration: security assessment, profile selection (Minimal / Standard / Hardened / Paranoid), tool detection for 20 security tools, auto-install missing dependencies, OS optimization, ML model setup.
Sentinel mode activates automatically. Your system is now being monitored. The AI HUD provides real-time interpretations in a bottom panel.
Natural language interaction. Ask questions, request scans, investigate threats. The AI selects the right tools, correlates findings, and builds a persistent knowledge graph.
Add honeypots. Set up bug hunting scope. Train your team in the Arena. Deploy the web dashboard for your SOC. Every feature unlocks deliberately through the safeguard system — never by accident.
nmap, Burp, Suricata, Zeek — they're each great at one thing. Dagda's aim is to orchestrate them through a single AI that keeps context across tools, correlates findings through a knowledge graph, and builds a persistent model of your environment as it goes.
SIEMs are built to collect and search, and they do it well. Dagda is exploring what comes after: Bayesian threat assessment, temporal correlation, and behavioral fingerprinting that feed into action — deploying honeypots, hardening systems, responding to incidents — all gated by the safeguard system.
Many AI security assistants wrap an LLM around a handful of API calls. Dagda's bet is that the hard part is everything around the model: hundreds of tools behind a Rust-enforced permission gate, scope-locked targeting, hash-chained audit logs, and 8 hard guardrails.
The core is AGPL-3.0 — meant to be read line by line, audited, forked, and improved. The source isn't public yet while it's in development.
1069 Python tests + 33 Rust tests. CI/CD with GitHub Actions. The safety systems get tested as hard as the features.
No data collection, no analytics, no phone-home — by design. With the Ollama local LLM it's meant to run air-gapped, so your data stays on your machine.
The release plan: signed releases with threshold signatures, and no auto-update. You verify before you deploy.
Dagda monitors its own integrity. SHA-256 baseline verification of all source files on startup, CycloneDX SBOM of its own dependencies, and a 10-point self-hardening audit. If critical files are tampered with, Dagda halts before loading.
Dagda is in development and not publicly deployed. The hard parts are the interesting ones: autonomous response you can trust, offensive tooling that stays inside its scope, and safeguards that hold up under pressure.