Deception Check / AI-tool file hunting / retained-data recount refreshed November 1, 2026 Window: September 8 through October 7, 2026, UTC. Unit: unique full exported request rows after fixed timestamp filtering and exact equality deduplication across the export's fields. 2,441 in-window exported rows became 2,317 after collapsing 124 exact duplicates. AI paths: 1,429; other selected paths: 888. Exact path /.claude/.credentials.json: 21, a subset of AI paths, not a third disjoint category. Historical AI classifier, case-insensitive substring search on the exported URL path: \.claude/|\.cursor/|mcp\.json|mcp_config|anthropic|openai|litellm|\.aider|\.continue/|windsurf|claude_desktop_config|codeium|huggingface|gemini Exact-path count uses case-sensitive equality without decoding. All selected rows in this export are GET. The extractor limits exported paths to 120 characters; no in-window exported path reached that limit. Status coverage is incomplete and response bodies are absent. Counts do not establish unique original requests, people, exposed files, credential use, or compromise. Exact duplicate collapse can combine otherwise indistinguishable exported rows. The detection pack's ten selected path endings are a practical monitoring subset. They are not the historical broad classifier and do not reproduce the total of 1,429. Extend monitoring to the tools actually used in your organization. Product configuration locations can change; requests for a name do not establish that a current product uses it. Raw operational data, source addresses, source identifiers, credential material, and fleet counts are not distributed.