Beacon embeds six watermark layers in every transcript, email, document, and meeting recording you send. Recipients hear the announcement. They see the markers exist. They don't know which marker is theirs. If it leaks, you can attribute it. If it doesn't leak, the deterrent worked.
Trusted recipients pasting transcripts into ChatGPT for a summary. Forwarding board memos to a vendor who feeds them to Copilot. Screenshotting documents into a group chat. Not malice. Just the friction-free workflow that AI now makes trivial — and the leak surface most security tooling ignores.
A meeting transcript carries every word said. One forward turns a confidential discussion into a search-indexable artefact. Without a per-recipient marker, you can't trace it back.
The board update copied into ChatGPT for a one-line summary travels to the model's training pipeline. The recipient thinks they've been efficient. You see nothing.
A meeting recording exported from Zoom or Otter, attached to an email, dropped in Slack — the bytes are identical for every recipient. Without per-recipient marking, the leaker is one of forty.
The watermark must look like ordinary content but uniquely identify the recipient when you know where to look. Six layers, each with documented defeat conditions. The system is honest about what it catches.
Per-recipient sentence rewrites at deterministic positions. Survives plain copy, screenshot+OCR, and light AI rewrite. Defeated by full rewrite from scratch.
Per-recipient style budget — punctuation preferences, sentence-length distribution, contraction frequency. Survives mild edits. Defeated by adversarial full rewrite.
Fingerprint-safe content variation — throwaway figures, immaterial names, neutral examples. Never load-bearing facts. Survives full re-typing. Defeated by cross-reference with another recipient's copy.
Specific neutral phrases at specific positions, varied per recipient. Lighter weight than the heavy layers, weaker against partial copy.
Zero-width characters encoding the recipient ID at sentence boundaries. Survives format-preserving copy. Defeated by plain-text paste or "show invisibles."
Document properties, email headers, font micro-variations. Last line of defence — strip-easy but cheap to add. Survives only same-format forwarding.
Every Beacon-protected meeting, email, or document opens with the same announcement. Recipients know the markers exist; they don't know which marker is theirs. Consent and deterrent in one move. Legal review owns the wording before any commercial use.
"This call is being recorded and transcribed. Each participant will receive a personalised version. Transcripts and recordings contain unique identifying markers; if your version appears outside agreed channels, the source can be identified."
The actual identifying markers are invisible to the recipient. They know the markers exist; they don't know what they are or where. The announcement does not name any specific token.
For meeting recordings, each recipient's greeting surfaces their unique trigger phrase prominently — a distinct, named, memorable per-recipient marker the recipient associates with themselves. The realisation is the deterrent.
Spread-spectrum audio watermarking under the audible greeting. Inaudible, unsearchable, survives Opus / MP3 / lossy compression. If the recipient edits the audible portion out, the DSSS payload still identifies them.
One watermarking engine, one registry, one investigation tool. Each surface composes the layers appropriate to its medium.
Per-recipient transcript and recording for every meeting. Audible greeting, hidden DSSS layer, continuous DSSS carrier across the call. The bot joins, admits, watermarks, delivers.
Per-recipient lexical, stylometric, and semantic variants. Subject untouched, body uniquely fingerprinted. Christie layers carry the load because format dies on forward.
Markdown, plain text, and Word (.docx) per-recipient variants. Per-paragraph invisible-character watermark with document structure preserved.
Paste a suspect snippet — text from a forward, transcribed audio, a chatbot reply. Beacon returns a ranked recipient list with per-layer scores. Human review required before any attribution.
Schedule periodic probes against ChatGPT, Claude, and other LLMs. Each registered canary phrase is queried; reproduction is scored and logged. Detection without waiting for the leak to surface in front of you.
Physical room device emitting an acoustic watermark embedded in any recording made in the room — including covert personal-device captures. Hardware product, separate development arc.
Four stages on every artefact. Every recipient registered. Every layer logged. The investigation surface is one CLI command — or one paste into the web interface.
The legal-reviewed announcement opens every meeting, every email thread, every document distribution. Consent record + deterrent in one move.
Layers compose at delivery time. Each recipient receives a unique artefact. The registry records every layer's fingerprint per recipient.
Hand each variant to your normal transport — Resend, Gmail, Zoom delivery, Notion share. Beacon is library- shaped; transport is your choice.
A leak surfaces. Paste the snippet. Beacon runs every matcher across every registered artefact, combines the scores, returns a ranked recipient list. Human review is the next step. Every time.
A 90-day pilot covers the announcement template (legal review included), the watermarking integration on your two highest-risk surfaces, and a confidential investigation surface for your team. Custom pricing based on artefact volume.
Human review required before any attribution.