Shadow AI Discovery · Open source

See every AI tool your team uses, before regulators ask.

Most companies have no idea which AI tools their employees actually use — or what data is being pasted into them. Floodlight gives you the inventory, the risk classification, and the regulatory mapping in seven days. Free.

No credit card. No backend account. The browser extension is open-source — you can read every byte of code that runs on your team's machines.

Real audit excerpt
Shadow AI Audit · Acme Capital Ltd
FL-2026-05
19
distinct AI tools observed
7
credential pastes flagged
ChatGPTHIGH
DeepSeekCRITICAL
GitHub CopilotMEDIUM
ClaudeLOW
The problem

You can't govern what you can't see.

DLP and CASB tools were not built for the modern AI taxonomy. They flag "ChatGPT" as one thing — when in reality there are five distinct ChatGPT tiers with five different data policies. The difference between them is the difference between a SOC 2-covered Enterprise contract and pasting customer PII into an account that trains on it.

i

23 tools, on average

A typical 100-person company uses around 23 distinct AI products across browser, desktop and embedded surfaces. IT knows about three or four. Security teams discover the rest after an incident.

ii

EU AI Act, in force.

Article 4 (AI literacy) is in force since 2 February 2025. As a deployer of any third-party AI system, you must take measures to ensure sufficient AI literacy among staff — and demonstrate them.

iii

Credentials in prompts.

API keys, customer data, source code — pasted into chat interfaces every day. Most companies don't find out until a leaked key triggers an alert in someone else's logs.

How it works

Seven days. One PDF. Zero hosted infrastructure.

The audit runs entirely on your team's own machines. Pasted content is classified locally — only category labels and counts ever leave the endpoint. The actual prompts never do.

Install the open-source extension

Sideload it on Chrome or Edge. The extension monitors visits to ~60 AI tool domains and classifies pasted content locally for emails, API keys, source code, regulated identifiers, and 12 other categories.

Use your browser as normal — for seven days

The extension runs silently. No popups, no prompts, no productivity drag. Metadata is logged on the device only — prompt content is never transmitted, and you can audit the source code yourself.

Receive a branded audit report

At the end of the period, export the local event log and we generate a 10–12 page PDF: tool inventory, risk classification, sensitive content events, regulatory mapping (EU AI Act, UK GDPR, FCA, ISO 42001), and recommendations.

What you receive

A document the CISO can read in fifteen minutes — and act on the same day.

Real example below. 11 pages, A4, mapped to specific regulatory obligations. Cites the actual usage data from the audit period — not a generic template.

  • §1
    Executive summary Headline narrative, hero stats, risk distribution, top three actions for this week.
  • §2
    Tool inventory Every AI tool observed, ranked by risk band. Event count and distinct user count per tool.
  • §3
    Sensitive content events Categories the local classifier flagged. Chronological list of high-concern pastes.
  • §4
    Regulatory mapping EU AI Act Articles 4 / 5 / 9–10–13. UK GDPR Articles 5(1)(f) / 32 / 44. FCA SYSC 6.1.1R / 13. ISO/IEC 42001.
  • §5
    Recommendations Eight actions ranked by impact-per-effort: this week, this month, quarterly. With time estimates.
  • §6
    Methodology & disclosure Privacy contract, limitations, "what this report is not". The trust artifact.
§1 · EXECUTIVE SUMMARY

The headline finding

19
distinct tools
28
active users
641
high-risk events
7
credential pastes
ChatGPT (Free/Plus)HIGH
DeepSeekCRITICAL
CursorMEDIUM
Get your own report — 7 days, no payment →
Open source & auditable

The thing security people actually trust.

We are asking your employees to install software on their work machines. That is a meaningful ask. The least we can do is publish every byte of it — the extension, the classifier, the risk scoring, the report template — so anyone can verify the privacy claims hold.

The privacy contract

Pasted content is classified locally, in the browser, and immediately discarded. Only category labels and counts leave the endpoint. The actual prompt text is never logged, never transmitted, never stored.

Read the byte-by-byte disclosure →

The risk methodology

Every tool in the taxonomy gets a numeric risk score and a band based on four published factors: trains-on-input default, hosting jurisdiction, compliance certifications, retention defaults. Scoring rules are public and corrections are welcome via PR.

Read the scoring methodology →

The full source code

Browser extension, classifier, audit report generator, taxonomy data — all of it lives in one public repository under MIT licence. Fork it, audit it, run it offline, send a pull request.

github.com/floodlightsecurity/floodlight →

What we don't do

We don't ship a closed-source agent. We don't require a hosted account to run an audit. We don't sell employee browsing data, ever, full stop. We don't pretend the extension catches mobile or BYOD usage — it doesn't, and the methodology section says so.

See the published limitations →
Free audit · No payment

Find out what your team is doing in the dark.

Sign up, install the extension on a few volunteer browsers, and in seven days you'll have the full PDF in your inbox. We'll be in touch within one working day.