One ambiguous prompt, one destructive command.
An AI agent can cheerfully run rm -rf, curl | sh, npm install <random> or aws s3 rm, sometimes from a single typo, sometimes from a malicious instruction buried in context the model just ingested.
Proud to be part of the OpenAI Partner Network and a registered member of the Claude Partner Network.
AI coding agent governance
Govern every command your AI coding agents run.
Run DarkControlAI in watch-only mode for 7 days. Claude Code, Codex, Cursor, Copilot and more. Up to 10 devices. No credit card. Nothing blocked.
Built for the AI coding agents your engineers actually use
See it in action
From shadow AI to governed AI in under 30 seconds.
Watch DarkControlAI intercept a Claude Code agent attempting destructive commands on a developer workstation, escalate the call to a human operator, and record the whole exchange, fleet-wide, in real time.
Install the agent
One-line installer. Windows, Linux, macOS. MDM-friendly.
Set your rules
Allow, deny or ask, by command, package, organisation or department.
Govern every command
Every attempt your AI coding agents make is matched against policy before it runs.
Prove it
Auditor-ready evidence on demand, centralised.
The Problem
Your AI coding agents can already run anything.
Claude Code, OpenCode, Codex, Copilot, Cursor, Google Antigravity, Grok Build and VS Code don't just suggest commands anymore; they execute them. With your developer's credentials. On your endpoints. Against your repositories. The controls to govern that haven't existed yet.
Per-machine settings don't scale.
Every AI coding tool ships its own permission file, configured per-laptop, per-developer, per-tool. There's no central place to say "engineering can install npm packages, but no one in finance can run apt." So nobody does.
No fleet-wide visibility. Anywhere.
Which machines run AI agents? Which agents tried what last week? Which were blocked, which slipped through? Most security teams genuinely cannot answer these questions today; there is no inventory and no log.
When something breaks, you have nothing.
If an AI agent does the wrong thing, security teams need an audit trail to investigate against. Static policy documents don't satisfy NIS2, ISO 27001 or the EU AI Act either; regulators want runtime evidence of human oversight.
The Solution
Allow. Deny. Ask. Audited.
DarkControlAI sits between your AI coding agents and your operating system. Every command they attempt is matched against the policy you set in a single console, and the verdict, plus a full audit record, lands on your security team's screen in real time.
ALLOW
Trust the safe stuff.
Read-only commands, repo navigation, test runs and your team's pre-vetted tooling clear instantly and silently. Developers feel nothing.
git status · npm test · python -m pytest
ASK
Put a human in the loop.
For commands that aren't quite blocked but warrant a glance, a package install, a system change; DarkControlAI escalates to a real person before the AI proceeds.
npm install <new package> · sudo apt install …
DENY
Stop the bad ones cold.
The dangerous patterns: destructive deletes, piped curls, package poisoning attempts, never reach the shell. The agent is told no, the event is logged, the machine is fine.
rm -rf /* · curl … | sh · aws s3 rm s3://prod
Central policy, fleet-wide reach
Set rules once in the console. Every endpoint picks them up automatically. No mass SSH, no per-machine drift.
Departments, not just orgs
Engineering can have a permissive ruleset. Finance can ban every package install. Contractors can have their own posture entirely.
Package-manager aware
One rule covers every install variant. npm install, npm i, yarn, pnpm, pip, apt, brew; DarkControlAI expands them for you.
Immutable audit trail
Every attempt: allowed, denied, asked, recorded with command, host, IP, service and timestamp. Rule-based alerts stream to your SIEM over syslog or HTTPS, and CSV export is built in.
Live compliance posture
A continuously-scored 0-100 view of license health, rule coverage, agent liveness and governance, with a 12-item checklist auditors recognise.
One agent, every OS
A single lightweight agent for Windows, Linux and macOS. One-line installer. Self-updating. Revoke any endpoint in one click.
What DarkControl does
Governance for every command your agents run.
Four capabilities, one control plane - from the instant an AI agent proposes a command to the audit trail your compliance team signs off on.
Real-time
See every command the instant it’s proposed
DarkControl sits inline with every AI coding agent your engineers use — Claude Code, OpenCode, Codex, Copilot, Cursor, Google Antigravity, Grok Build and VS Code. Every shell command, file write and network call is judged against your policy before it can run - in single-digit milliseconds.
- Inline verdicts - allow, ask or deny on every single action.
- Zero agent changes - drop-in, works with the tools your engineers already use.
- Full context - command, working directory, agent identity and repo, all captured.
Policy
Write a rule once. Enforce it everywhere.
Express intent in plain glob and regex patterns. Rules cascade from organisation to department to group, so a single deny on destructive commands protects every project at once.
- Layered policy - org, department and group rules merge predictably.
- Safe by default - anything unmatched falls back to “ask”.
- Instant rollout - new rules apply on the next command, no redeploy.
Insight
Know what your agents are doing - at a glance
Every verdict streams into a live dashboard. Spot the noisiest repositories, the most-blocked commands and unusual spikes before they ever become incidents.
- Live metrics - verdicts, agents and blocked actions in real time.
- Trends - 14-day activity, top commands and per-team breakdowns.
- SIEM alerting - rule-based alerts over syslog or HTTPS, plus CSV export.
Assurance
Coverage you can prove to an auditor
DarkControl governs every major AI coding agent and keeps an immutable, time-stamped record of every decision - mapped to the frameworks your auditors already ask about.
- All major agents - one policy across your whole toolchain.
- Immutable audit log - tamper-evident and retained for as long as you need.
- Framework-aligned - EU AI Act, NIS2, ISO 27001 and SOC 2.
CCClaude CodeMonitored
CUCursorMonitored
CXCodexMonitored
OCOpenCodeMonitored
VSVS CodeMonitored
GCGitHub CopilotMonitored
AGGoogle AntigravityMonitored
GRGrok BuildMonitored
EU AI Act
NIS2
ISO 27001
SOC 2
Beyond enforcement
Context, cost and evidence around every verdict.
Enforcement is where DarkControl starts. These capabilities turn each decision into a complete session record, a cost signal and audit evidence your security and compliance teams can act on.
Agent sessions
Follow any agent run end to end. Turns, tool calls and verdicts are grouped into the session they came from using each tool’s own session ID, with subagents separated from the main agent and every event classified.
Token usage
Input and output tokens per session and per AI tool, read from each tool’s own usage records rather than estimated. See what AI-assisted engineering actually consumes, and where.
AI Analysis Enterprise
AI-driven review of your telemetry: anomalies, rule recommendations backtested against your own history, rule hygiene and agent health, summarised in a concise digest. Runs on DarkControl’s own GPU infrastructure; your data is never sent to a third-party AI provider.
SIEM alerting
Define alert rules on the events that matter and deliver them to your SIEM over syslog (UDP or TCP) or HTTPS. Thresholds, time windows, cooldowns, field selection, on-demand test alerts and a full firing history.
EU AI Act module
Maps the deployer obligations of the EU AI Act to the controls DarkControl already enforces, scored alongside SOC 2, CIS and ISO 27001, with an attestation register for the organisational obligations. It supports your compliance assessment; it does not certify conformity.
Named policies with precedence
Build reusable, named policies and assign them to an organisation, department, endpoint group or individual endpoint. A precedence engine resolves them into one effective ruleset per device: endpoint, then group, then department, then organisation.
Microsoft Entra ID SSO
Console sign-in through your Microsoft Entra ID tenant, so access to DarkControl follows the identity lifecycle you already manage.
Privacy by design
Metadata-only capture by default, with secrets redacted on the endpoint before anything is sent. Full prompt capture is a deliberate, visible decision that only the platform administrator can enable.
Get in touch
See DarkControlAI against your own AI tools.
Tell us what your engineers run and where, and we'll set up a private demo against a representative environment. A real human from DarkControl will reply within one business day.