All engagements conducted under NDA·Evidence preserved to technical chain-of-custody standards·Engagements limited to qualified organizations

ServicesThe EvidenceApproachAboutResource LibraryResearchRequest a Confidential Briefing
Request Briefing

Three Time Bombs Every Company Is Sitting On

Before AI governance policies existed, your employees were already using AI tools. That data is gone—or is it?

Time Bomb 01

Regulatory Violations

  • HIPAA violations from PHI fed into ChatGPT — penalties up to $2,190,294 per violation category per year (effective Jan 28, 2026)
  • SOX internal control gaps when material nonpublic information enters AI tools
  • Attorney-client privilege waived when client matter details are processed through a consumer AI account outside privilege protections
Time Bomb 02

Litigation Exposure

  • Plaintiffs subpoena AI vendor records (OpenAI, Anthropic) to prove negligence
  • Without forensic documentation, you have no evidence to present when challenged
  • Insurance carriers now denying claims without AI usage documentation
Time Bomb 03

Trade Secret Contamination

  • Proprietary algorithms, code, and processes fed into AI tools
  • May have "escaped" into model training data permanently
  • Destroys IP protection and competitive advantage

Three Artifacts That Survive Deletion

Documented in our technical reference and verifiable on any Windows endpoint running Chrome. Each recoverable through standard tooling — no specialist hardware required.

01

IndexedDB Write-Ahead Log

Chromium browsers store AI tool conversation data in a LevelDB write-ahead log. Deleting a conversation appends a tombstone marker — the original content remains at its prior offset until compaction. Recoverable as plain text with standard tooling.

02

Browser History (SQLite)

Chrome and Edge record every AI tool domain visited, with timestamps and visit counts, in a SQLite database. Clearing browser history does not guarantee the underlying records are overwritten.

03

Session Restore Files

Chrome serialises open tab state to disk at browser close for crash recovery. Where an AI tool session was open, the captured state can include conversation content — independent of the history database and the IndexedDB log.

Full technical detail, file paths, and a reproducible verification test: Read the technical reference →

What This Methodology Is Built to Surface

What the audit process is designed to recover from a standard Windows endpoint running Chromium-based browsers, mapped to common enterprise deployment scenarios.

The Scenario

A mid-size regulated organization. Existing DLP solution in place. An acceptable use policy either drafted or in progress. IT has blocked a handful of known AI domains on the corporate network. Leadership considers the AI governance situation handled. From the outside — and from their own audit logs — everything looks fine.

Finding 01

Prompt Content Surviving Interface Deletion

Browser artifact analysis is designed to recover content from AI sessions — including from tools the organisation never approved. The IndexedDB write-ahead log retains the original conversation text after interface deletion, because the deletion appends a marker rather than removing data. This is what DLP does not flag and what survives browser history clearing, because the artifact file is entirely separate from browser history.

Finding 02

OAuth Tokens for Unapproved Platforms

Authentication tokens for AI platforms outside the organisation's approved vendor list may be present in browser credential storage on endpoints where those tools were accessed. These tokens grant ongoing access to platforms where prior sessions — and in some cases, uploaded documents — may still be accessible. Where the employee no longer works at the organisation, the token may still be valid.

Finding 03

Conversation Exports Outside DLP Scope

AI platform conversation exports saved to user download folders sit outside any monitored directory and outside DLP coverage. Where regulated data was processed in the session, the export contains the full text. These files are named generically and are not distinguished from other downloads by endpoint monitoring tools.

Finding 04

Browser Extension Access to Webmail and Editors

AI writing assistant browser extensions with granted permissions to read and modify content across all sites — including webmail, document editors, and internal portals. These extensions operate outside endpoint DLP, outside network monitoring, and outside the organization's AI tool inventory entirely. Permission was granted at installation. Nobody reviewed it.

What This Means Legally

Under HIPAA, every cached prompt fragment containing patient information is a potential impermissible disclosure — regardless of whether the employee intended to expose it. Under SOX, AI-assisted financial analysis not archived in enterprise systems creates a documentation gap regulators are increasingly equipped to exploit. Under attorney-client privilege rules, client matter details processed through a consumer AI account may constitute a waiver.

The finding that surprises most compliance teams: the employees were not being malicious. They were being efficient. The tools were useful and fast and nobody told them the session didn't end when they closed the tab.

Scenarios above illustrate what the methodology is designed to surface. Actual recoverable artifacts depend on endpoint configuration, AI tools in use, browser version, and time elapsed since last session. No two endpoints are identical.

Request a Confidential Briefing