Endpoint Security for the AI Era

Know what AI agents actually do on your machines.

AI coding assistants run code on Windows workstations with almost no oversight. AI Trace gives security teams a complete record of what each one actually did: every program it launched, every file it touched, every network connection it made, all tied back to the specific AI session that caused it.

Request Early Access Learn More

The Problem

AI agents operate in a blind spot

AI coding assistants start programs, write files, and reach out to the internet on workstations. Traditional endpoint security tools weren't built to link those actions back to the AI agent that caused them, or to present them in a form a reviewer can sign off on.

Untracked Execution

Arbitrary shell commands

AI agents start PowerShell, cmd, and other command-line programs to run the code they generate. These child programs have the same access as the user running them, and standard logs don't record which AI kicked them off.

Data Exposure

Sensitive file access

Nothing stops an AI agent from reading SSH keys, cloud credentials, auth tokens, or any file the user has access to. Traditional endpoint security won't flag it, because the user account is authorised to read those files.

Exfiltration Risk

Unrestricted network access

Programs started by an AI agent can reach the internet freely: make web requests, look up any domain, open connections to any server. Windows itself doesn't restrict or log any of it.

No Audit Trail

Nothing to review after the fact

When an AI session ends, the chat transcript shows what the user asked and the end result on disk, but not what happened in between. Which files were opened, what commands ran, where traffic went: none of it is recorded. Incident-response teams have nothing to work with.

Supply Chain

Untrusted code execution

Code the AI generates runs immediately on the workstation. Dependencies get downloaded and installed without review. Build scripts run with the user's full access; nothing is sandboxed.

No Attribution

Invisible agent provenance

When an AI agent starts a program that starts another program, the chain of who-started-what gets lost. Your security team can't tell whether a PowerShell window was opened by the user or by an AI agent three steps removed from it.


Capabilities

Full-stack agent telemetry

A lightweight Windows service plugs into the operating system's own kernel-level event stream (Microsoft's ETW), the same low-level telemetry the OS uses to track what every running program does. Nothing is injected into other processes, so there's nothing for an AI agent to bypass and nothing for an attacker to quietly disable. Every captured event (programs launching, files being touched, network connections, registry changes that plant persistence) is tagged with the AI agent session that caused it.

CORE

Deep Event Capture

Records every relevant action (programs starting and stopping, files opened, written, renamed or deleted, network and website activity, and system-configuration changes), read directly from inside Windows. Tamper-resistant by design: no software is injected into other programs, and there are no hooks an attacker can quietly remove.

TREE

AI Session Attribution

Automatically identifies running AI coding assistants and tracks every program they start, plus every program those programs start, all the way down. Every action links back to the specific AI session that triggered the chain.

NET

Network Telemetry

Logs every outbound connection and every website lookup made by an AI agent or anything it launched. A built-in watchlist flags connections to paste sites, anonymous file drops, and URL shorteners, destinations an AI coding assistant has no business reaching. Uncommon connection types (like direct traffic to unfamiliar servers outside normal web channels) are surfaced separately, since they're the channels attackers often use for covert control.

FS

File System Monitoring

Tracks every file an AI agent or anything it launched opens, writes, renames, or deletes. Built-in rules flag touches to SSH keys, cloud credentials, Kubernetes config, .env files, and browser secrets, and security teams can edit the rules to match their own definitions of sensitive.

CMD

Command & Script Analysis

Records the full command line of every program an AI agent launches, and scans for known attack patterns: obfuscated PowerShell, download-and-run one-liners, and the standard techniques for installing malware so it survives a reboot. Obfuscated commands are automatically decoded so reviewers see the real script, not a meaningless blob.

PERSIST

Persistence Detection

Flags the specific Windows system-configuration changes attackers use to make code survive a reboot: auto-start entries, service installs, scheduled tasks, and the other textbook persistence tricks. An AI coding assistant has no legitimate reason to make any of these changes, so each one is surfaced with a clear label of which trick was used.

DROP

Dropper Detection

Catches the classic attack pattern where one program writes an executable file to disk and another program runs it moments later. The write and the run are correlated across the whole AI session, and each alert explains exactly why it was flagged, not just that something looks off.

SESSION

Session Review Workflow

Every AI session becomes its own reviewable record, bundled from the moment the agent started through the moment it finished. The dashboard groups them into a needs-review queue, assigns a plain severity label (critical, high, medium, low, clean), lets reviewers sign off one click at a time, and keeps an audit log of who reviewed what and when.

PDF

Forensic Session Reports

Every session has a printable report with the full event list, the severity verdict, and the tracking details that make it defensible as evidence: session ID, when it was captured, when the report was produced. Save-as-PDF gives you a standalone document for compliance hand-off or post-incident review.

HOOK

Webhook Alerts

The moment an AI session crosses the risk threshold, your team gets notified, delivered to Slack, Discord, Mattermost, Google Chat, or any tool that accepts a standard webhook. Structured data feeds are also available if you want to pipe everything into an existing SIEM or logging system.

POLICY

Configurable Rules & Retention

Security teams can add or remove watchlist entries (sensitive file paths, suspicious hostnames), adjust how often data is captured and how much of it, and set how long records are kept. Old data is aged out automatically, with a one-click purge for incident-response situations.


How It Works

From raw event to actionable alert

AI Trace runs as a small background service on each Windows workstation. It watches activity in real time, figures out which actions belong to which AI agent, and sends the results to a central dashboard your security team reviews.

01

Agent Detection

Identifies running AI coding assistants by program name, install path, and behaviour. Keeps a live list of active sessions and the programs they started from.

02

Activity Capture

Captures programs starting, files being touched, network and website activity, and system-configuration changes as they happen, read directly from inside Windows. Filtering happens on the workstation itself, so only activity tied to an AI agent ever leaves the machine.

03

Session Linking

Every event is linked back to the AI session that caused it, along with the full chain of parent-and-child programs. Related events are grouped into the sequences a reviewer can actually make sense of.

04

Rule Evaluation

Each session is scored against a set of rules: sensitive files touched, suspicious hosts contacted, suspicious commands run, shells launched inside an AI session, attempts to install persistence, and code loaded from files the agent itself wrote. Sessions come out tagged critical, high, medium, low, or clean.

05

Review & Alert

Flagged sessions land in the security team's needs-review queue, and a webhook fires once so the team sees each incident the moment it happens. Every session has a printable report, and the mark-reviewed workflow records which reviewer signed off and when.


Event Coverage

Comprehensive event capture

AI Trace captures the events listed below, every one tied to the AI session that caused it. The capture is deliberately narrow and honest: only metadata, never the contents of your files or the payload of network traffic.

Program start & end
Full command-line arguments
Parent / child program relationship
Full process chain per session
File opened (read or write)
File written
File renamed
File deleted
Sensitive-file alerts
Outbound network connections
Uncommon outbound channels (UDP)
Website name lookups (DNS)
Suspicious-host alerts
Suspicious-command detection
Obfuscated PowerShell decoded
Shell launches inside AI sessions
Persistence attempts flagged
Dropper-pattern detection
Severity label per session
Needs-review queue & sign-off
Forensic PDF session reports
Webhook alert when flagged
Data retention & on-demand purge
On-disk buffer if server is offline

Compatibility

Built for the agents your teams use

AI Trace ships with built-in detection for the AI agents below. The detection policy is a configuration file; adding a new agent is a config change, not a product release.

Claude Code
Anthropic
Claude Desktop
Anthropic
Cursor
Anysphere
Windsurf
Codeium
Codex CLI
OpenAI
Codex App
OpenAI
GitHub Copilot
GitHub
Gemini CLI
Google
Kiro CLI
AWS (formerly Amazon Q)
Cline
Open Source
Aider
Open Source
Custom Agents
Configurable signatures

Ship AI-assisted code with confidence.

AI Trace is in private early access. If your organisation uses AI coding assistants on Windows workstations and needs endpoint-level visibility into what they actually do, get in touch.

Request Early Access