GALAHAD
/Waterline
MechanismHow It WorksUse CasesGalahad HomeRequest Audit →
Galahad Labs · Forensic AI Data Exposure

Your confidential documents
surface in AI answers.

The board pack pasted into ChatGPT for "quick summary." The legal memo forwarded to a vendor who fed it to Copilot. The transcript a departing employee dropped into Otter. Waterline plants canaries in your documents, probes AI systems for reproduction, and tells you who forwarded what — with audit-grade evidence and a human review gate at every step.

Request a Waterline Audit →See the Mechanism
Waterline
ChatGPT, Claude, Gemini, Copilot, Perplexity — every confidential document forwarded to one of these is a candidate for leak. Waterline tells you which ones did.
The Surface

Most leaks aren't malicious.
They're convenient.

The leak surface security tooling ignores: a trusted recipient pasting confidential content into an LLM for a quick summary, forwarding a transcript on WhatsApp, or screenshotting a document into a group chat. Not malice. Just the friction-free workflow that AI and messaging apps now make trivial.

📄

Documents pasted into chatbots

A board pack copied into ChatGPT for a one-line summary travels to OpenAI's training pipeline. The user thinks they've been efficient. You can't see it happen.

📝

Transcripts forwarded sideways

Otter, Granola, and Fathom transcripts carry the full content of confidential calls. One forward turns a private meeting into a search-indexable artefact.

🔍

Detection arrives too late

You only learn a document leaked when an LLM regurgitates it back to a third party. Without canaries, you can't prove where it came from. Without probes, you don't see it until it's public.

Mechanism

Plant. Probe.
Detect. Takedown.

Four stages. Every artefact is registered. Every probe is logged. Every detection ends with the human-review gate.

Plant

Embed canaries in confidential docs & register

→
Probe

Query AI systems for canary reproduction

→
Detect

Match suspect text or audio back to the recipient

→
Takedown

Generate formal notices with evidence attached

🆕

Local-Compute Canaries

Document content never leaves the machine. Generation, embedding, registry, and detection all run locally. External calls only at the probe step, by design.

📚

Layered Watermarking

Six watermark layers ranked by survival: lexical paraphrase, stylometric signature, semantic fingerprint, steganographic placement, invisible characters, metadata. Survives copy-paste, screenshot-OCR, and light AI rewrite.

🔎

Multi-Target Probes

Probe Anthropic, OpenAI, and local models from one CLI. Probes ask the target to acknowledge familiarity, not reproduce content — reducing re-leak risk.

🧮

Confidence-Scored Attribution

Each detected leak carries per-layer match scores combined into a single confidence percentage. The threshold is yours; the human review is mandatory.

📜

Takedown-Ready Evidence

Export registry rows, probe transcripts, and match traces as a signed evidence pack. Ready to attach to formal takedown notices, regulatory filings, or expert-witness reports.

🔑

Custom Hotwords

Improve detection accuracy on domain-specific terms — names, technical jargon, programme codes. Hotwords seed the probe templates and the matcher in one configuration step.

How It Works

Audit-grade
from first plant.

Waterline runs as a CLI on your laptop or in your secured compute environment. Every command produces a registered artefact and a manifest. Nothing is reconstructed; everything is replayable.

01

Plant a canary

Run waterline plant document.pdf against any confidential document. The CLI generates per-recipient canaries, embeds them, and writes a registry row keyed on document and recipient.

02

Probe AI systems

waterline probe --target anthropic:claude-sonnetqueries the AI system with prompts derived from your registry. Responses are scored against the canary library and recorded in the probe results table.

03

Detect from suspect output

When a leak surfaces — a screenshot, a forwarded email, a chatbot reply — paste the text into waterline detect. Per-layer matches combine into a ranked-candidate list with confidence scores. Human review required before any attribution.

04

Generate the takedown

waterline takedown produces a formal notice with the detection evidence, registry entries, and match chain attached. Suitable for legal, regulatory, or contractual escalation.

Use Cases

Built for the documents
you can't afford to lose.

Board materials. M&A diligence rooms. Legal memos. Compliance files. Anything where attribution matters more than aesthetics.

👔

Board & investor materials

Plant per-recipient canaries before distribution. If the deck surfaces in an AI response or an unauthorised forward, you know which director it came from.

⚖️

Legal & regulatory

Privileged memos, expert reports, and filings carry forensic fingerprints. Probe results and detection traces attach directly to formal takedown notices.

💼

M&A & transactions

Diligence rooms see hundreds of recipients across counsel, bidders, and advisers. Per-recipient canaries make leak attribution a CLI command, not a forensic investigation.

Get Started

Find out what AI
already knows about your documents.

A 90-day audit covers planting, probing, and a confidential detection report on your highest-sensitivity document set. We run the audit; you keep the registry. Custom pricing based on document footprint and probe volume.

Request a Waterline Audit →See Beacon (per-recipient watermarking)

Human review required before any attribution.

Galahad
AI that knows its place. · Founded by Ross Barnes
hello@galahadgroup.co.uk

HomeServicesArticlesikigAIEnableGrailContact
© 2026 Galahad. All rights reserved.London · Global