v0.4.0 · Open Source · AGPL-3.0

Leave no
model untraced.

Forensically sound AI artifact collector & analyzer. Collect evidence from 26 AI platforms, map to MITRE ATLAS, generate STIX 2.1 reports. Built for DFIR professionals.

Quick Start View on GitHub
0
Collectors
0
Velo Artifacts
0
Report Formats
TRACE
trace — forensic session
$ trace discover
Discovered 5 AI platform(s):
• ollama (inference)
• hermes (agent)
• text_generation_webui (inference)
• llama_cpp (inference)
• huggingface (cloud)
$ trace collect --output /evidence --deep
✓ ollama: 14 artifacts collected
✓ hermes: 60 artifacts collected
✓ huggingface: 68 artifacts collected
✓ text_generation_webui: 2 artifacts
Collection complete: 144 artifacts from 4 platforms
$ trace analyze /evidence --mitre-atlas --risk-score
Timeline: 144 events
IOCs: 8,261 indicators found
ATLAS mappings: 2,235 technique mappings
Risk Score: 12/100 (Low)
$ trace report /evidence --format all
html: /evidence/report.html
json: /evidence/report.json
stix: /evidence/report.stix.json
$

Built for security professionals

Forensic Soundness Chain of Custody STIX 2.1 Export Cross-Platform Velociraptor Integration
Capabilities

Forensic-grade AI evidence collection

From discovery to courtroom-ready reports. Every step verified, every artifact hashed, every timestamp UTC.

Discover

Scan endpoints for 26 AI platforms — inference engines, agent frameworks, dev tools, cloud caches, Docker AI, browser AI, and live network AI traffic. One command reveals what's running.

Collect

Read-only forensic collection with SHA-256 hashing per file. Deep mode captures session history, conversations, and credentials.

Analyze

Unified timeline, IOC extraction (URLs, IPs, domains, API keys), MITRE ATLAS mapping, and risk scoring 0–100.

Report

Generate HTML forensic reports, structured JSON, and STIX 2.1 bundles. Ingest directly into MISP, OpenCTI, or your SIEM.

Fleet Deployment

15 Velociraptor VQL artifacts deploy across your fleet. Hunt for AI tools on every endpoint from a single server.

Forensically Sound

Read-only collection. SHA-256 hash verification. Chain of custody manifest. UTC timestamps. No agents, no footprints.

Go Binary — no Python required
Single Self-Contained Binary

Run TRACE without Python

Prebuilt executables for macOS, Linux, and Windows (amd64 + arm64). Stdlib-only Go — cross-compiles cleanly. One-shot sweep with run: discover → deep collect → HTML + JSON reports, with full chain of custody.

TRACE-branded CLI · live spinners · severity-colored output · system info in every scan
$ make -C go all
# builds bin/trace-{darwin,linux,windows}-{amd64,arm64}
$ ./bin/trace-darwin-arm64 run -o /tmp/evidence
✓ one-shot sweep → HTML + JSON + custody
$ ./bin/trace-darwin-arm64 scan
✓ quick risk summary (no files written)
26 Collectors

Every major AI platform

Inference engines, agent frameworks, development tools, and cloud caches — TRACE finds them all.

🦙

Ollama

14+ artifacts

Config, model manifests, Ed25519 signing keys, conversation DB, CLI history

🔬

LM Studio

8+ artifacts

Settings, LevelDB conversations, session store, model registry

🧠

GPT4All

6+ artifacts

chat.db (SQLite), settings.json, model cache paths

💬

text-generation-webui

6+ artifacts

settings.yaml, chat logs, character definitions

⚡

llama.cpp

2+ artifacts

Process detection, shell history, HuggingFace cache

🐉

KoboldCpp

4+ artifacts

Config JSON, session saves, process detection

🎓

Unsloth

4+ artifacts

Config, model cache, training logs, HF integration

🔀

LiteLLM

5+ artifacts

config.yaml, proxy logs, API key store, spend tracking

🌉

Bifrost

3+ artifacts

Gateway config, session cache, routing rules

🤖

Hermes

60+ artifacts

Sessions, state.db, memories, cron, secrets, skills, logs, auth

🔄

AutoGPT

4+ artifacts

ai_settings.yaml, .env, workspace, file_logger

👥

CrewAI

4+ artifacts

crewai.toml, .env, ChromaDB memory, knowledge base

🧑‍💻

Devin

5+ artifacts

Session state, workspace, auth tokens, command history

🧿

Eigent

4+ artifacts

Agent config, session logs, tool registry, credentials

🕵️

Shadow AI

meta-collector

Meta-collector: correlates evidence across all other collectors

🐳

Docker AI

3+ artifacts

Gordon assistant, AI model registry, running AI containers/images (ollama, vLLM, LocalAI, OpenWebUI)

✏️

Aider

3+ artifacts

.aider.chat.history.md, input history, tags cache

🎯

Cursor

4+ artifacts

globalStorage SQLite, .cursorrules, settings

🔮

Claude Code

4+ artifacts

~/.claude/ directory, projects, auth tokens

🐚

Shell-GPT

3+ artifacts

History, .sgptrc config, role definitions

🪐

Antigravity

4+ artifacts

Project state, session history, auth tokens

⬢

VSCodium

3+ artifacts

globalStorage SQLite, settings, extension state

🔎

Code Scanner

3+ artifacts

AI framework imports, MCP server configs, hardcoded API keys in source

🤗

HuggingFace

12+ artifacts

Model configs, refs, snapshots, auth tokens

🌐

Browser AI

3+ artifacts

ChatGPT, Claude, Gemini, Copilot, Perplexity, Brave Leo browser history & per-site stores

📡

Network AI

live

Live process→domain AI traffic classification against 100+ AI providers

Workflow

Discover → Collect → Analyze → Report

🔍
STEP 1

Discover

Detect installed AI platforms across the system — inference engines, agents, dev tools.

📦
STEP 2

Collect

Read-only forensic collection with SHA-256 hashes and chain of custody manifest.

📊
STEP 3

Analyze

IOC extraction, MITRE ATLAS mapping, unified timeline, risk scoring 0–100.

📋
STEP 4

Report

HTML forensic reports, structured JSON, STIX 2.1 bundles for MISP/OpenCTI.

Velociraptor Integration

Fleet-wide AI discovery

Deploy 15 Velociraptor VQL artifacts across your entire fleet. Hunt for AI tools on every endpoint from a single server — no agent installation required.

TRACE.AI.Inference
Ollama, LM Studio, GPT4All, text-generation-webui, llama.cpp, KoboldCpp
TRACE.AI.Agents
Hermes, AutoGPT, CrewAI, Aider, Shell-GPT
TRACE.AI.DevTools
Cursor, Claude Code, Continue, Cline, Warp
TRACE.AI.APIKeys + Network + NetworkAI + Processes + HuggingFace
Credential scanner, network detection, live AI traffic, process detection, HF cache
TRACE.AI.Browser + Docker + CodeScanner + ShadowAI
Browser AI history, Docker/Gordon workloads, source-code AI scanning, shadow-AI correlation
TRACE.AI.Binary.Linux · macOS · Windows
Download and run the TRACE Go binary on Linux, macOS, and Windows endpoints
Velociraptor VQL
# Hunt for AI inference across all endpoints
SELECT Hostname, Platform, Version,
ConfigPath, ModelCount
FROM Artifact(
TRACE.AI.Inference(
CollectChatHistory=TRUE,
DeepCollection=FALSE
)
)
# Find endpoints running Ollama
SELECT * FROM Artifact(
TRACE.AI.Processes)
WHERE ProcessName =~ "ollama"
Get Started in 30 Seconds

Install. Discover. Trace.

Install
$ pip install ionsec-trace
Discover
$ trace discover
Collect
$ trace collect -o /ev
CLI Reference
$ trace --help
Usage: trace [OPTIONS] COMMAND
run One-shot: discover → deep collect → HTML+JSON reports
discover Detect installed AI platforms
collect Collect forensic artifacts to directory
analyze Analyze evidence (timeline, IOCs, ATLAS, risk)
report Generate HTML/JSON/STIX reports
scan Quick triage scan
iris Push evidence into a DFIR-IRIS case
# Options:
--deep Include session-level data
--platforms Filter: ollama,hermes,cursor
--mitre-atlasMap findings to ATLAS techniques
--mitre-attackMap findings to MITRE ATT&CK
--risk-score Calculate risk score (0-100)
--format html | json | stix | all
Threat Intelligence

MITRE ATLAS mapping

TRACE automatically maps collected evidence to MITRE ATLAS techniques, giving you actionable threat intelligence.

AML.T0048
AI Tool Integration
Detect AI tool usage and configuration on endpoints
AML.T0025
Modify Model
Identify model tampering and unauthorized modifications
AML.T0050
LLM Data Exfiltration
Detect outbound data patterns from AI conversations
AML.T0055
Credential Access
Find exposed API keys and auth tokens in configs
12 out of 100 LOW HIGH

Risk scoring that matters

TRACE calculates a composite risk score across four categories, each weighted 0–25, giving a total of 0–100.

Credentials 0/25

Exposed API keys, auth tokens, .env files

Exfiltration 10/25

URLs/domains in conversations, outbound data patterns

Jailbreak 0/25

Prompt injection patterns, system prompt leakage

Autonomy 2/25

Agent frameworks, autonomous execution evidence

FAQ

Frequently asked questions

What does "forensically sound" mean?
Read-only collection (no source modification), SHA-256 hash verification per file, chain of custody manifest with timestamps and tool version, all timestamps in UTC, and evidence packages are append-only. TRACE leaves zero footprint on the target system.
What platforms does TRACE support?
26 collectors across 6 detection layers: Inference engines (Ollama, LM Studio, GPT4All, text-generation-webui, llama.cpp, KoboldCpp, Unsloth, LiteLLM, Bifrost), Agent frameworks (Hermes, AutoGPT, CrewAI, Aider, Shell-GPT, Devin, Eigent, Shadow AI), Dev tools (Cursor, Claude Code, Antigravity, VSCodium, Code Scanner), Cloud caches (HuggingFace), Docker AI (Gordon + hosted LLM workloads), Browser AI (Brave Leo, Perplexity, Copilot, ChatGPT, Gemini), and live Network AI traffic detection. A Shadow AI meta-collector correlates evidence across all collectors.
How does Velociraptor integration work?
TRACE ships 15 Velociraptor VQL artifact YAML files. Upload them to your Velociraptor server and deploy across your fleet. No agent installation needed — Velociraptor handles remote collection. Artifacts cover Inference, Agents, DevTools, APIKeys, HuggingFace, Network, NetworkAI, Processes, Browser, Docker, CodeScanner, ShadowAI, and the Binary artifacts (TRACE.AI.Binary.Linux, TRACE.AI.Binary.macOS, TRACE.AI.Binary.Windows) that download and run the TRACE Go binary on Linux, macOS, and Windows endpoints.
Can I use TRACE in incident response?
Absolutely. TRACE was designed for DFIR professionals. The chain of custody manifest, SHA-256 hashes, and UTC timestamps make evidence court-ready. STIX 2.1 output integrates with MISP, OpenCTI, and major SIEM platforms.
What license is TRACE under?
TRACE is released under AGPL-3.0-or-later. Free to use, modify, and distribute. Contributions welcome via GitHub.

Ready to trace?

Install in 30 seconds. Discover AI platforms. Collect evidence. Generate reports. All forensically sound.

View on GitHub
$ pip install ionsec-trace