Solution Paper

VPNs and TLS Protect the Pipe, Not What AI Providers Keep: Sources

The sources behind VPNs and TLS Protect the Pipe, Not What AI Providers Keep, numbered as in the paper.

Provider terms, prices and product names change. Follow the links for the current wording.

  1. Internet Engineering Task Force, “The Transport Layer Security (TLS) Protocol Version 1.3”, RFC 8446, Section 1, confidentiality clause “data sent over the channel after establishment is only visible to the endpoints”
  2. OpenAI, “Data controls in the OpenAI platform”, the API data guide, which carries no revision date of its own, so its statements are dated only by the read date; what it states is that abuse monitoring logs are generated for all API feature usage and retained for up to 30 days and that those logs may contain prompts, responses and derived metadata, which is narrower than retaining inputs and outputs
  3. Anthropic Privacy Center, “How long do you store my organization’s data?”, page dated July 1, 2026: 30 days on the API, inputs and outputs retained up to 2 years where flagged by trust and safety, and User Safety classifier scores up to 7 years
  4. Preservation order in New York Times v. OpenAI, Southern District of New York, entered 2025-05-13 by Magistrate Judge Ona T. Wang, ordering output log data preserved “regardless of whether deletion was requested by users or mandated by privacy regulations”, with enterprise and education agreements carved out. The order itself was not read; the wording comes from legal commentary, National Law Review, “Privacy Under Pressure: What NYT v. OpenAI Teaches Us About Data Governance”, a title taken from the URL slug because no separate title string was recorded (not verified from a primary source)
  5. Microsoft, “Microsoft Product and Services Data Protection Addendum”, May 2026 edition, self-dated 05/22/2026 for the clarification of subprocessors supporting artificial intelligence functionality: six months’ notice for a new customer-data subprocessor, 30 days for AI-functionality subprocessors, and a right to disable that subprocessor until at least six months after notice, where the paper drops “at least”. The document was retrieved as MicrosoftProductandServicesDPA(WW)(English)(May2026)(CR).docx at HTTP 200, 6,510,568 bytes, and extracted to text. One limit worth carrying: the addendum itself was read in full, while the vendor pages listing the individual artificial-intelligence subprocessors were not retrievable to us
  6. FedRAMP Marketplace product listings, read from the live server-rendered site rather than from vendor claims: the deprecated marketplace.fedramp.gov is an empty shell that redirects, the terminology is now FedRAMP Certified rather than Authorized, listings distinguish Certified Services inside the assessment scope from Other Service(s) in the system but outside it, and the pages carry no last-updated date of their own
  7. Hornetz, Yavarzadeh (joint first), Cheu, Gascon, Gerlach, Moghimi, Schoppmann, Schwarz and Zhang, “TDXRay: Microarchitectural Side-Channel Analysis of Intel TDX for Real-World Workloads”, UC San Diego, CISPA and Google: cache-line-granular memory access traces of unmodified confidential VMs, recovering user prompts from a single trace during tokenization, with an ethics statement that the work fully complies with the TDX security model. The PDF was opened; the venue is inferred from the PDF filename and a silicon vendor’s bulletin date and is therefore not stated here as fact
  8. Nasr, Carlini, Sitawarin, Schulhoff, Hayes, Ilie, Pluto, Song, Chaudhari, Shumailov, Thakurta, Xiao, Terzis and Tramèr, “The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt Injections”, arXiv:2510.09023, submitted 10 October 2025: twelve recent defenses bypassed with attack success above 90 percent, most of which had reported near-zero
  9. GitHub Terms of Service section J, the terms behind section 11’s license-tier point: Input is “any content or context provided to an AI Feature, including prompts”, plus “attachments, code in your workspace, and conversation history”, and J.3 grants “a license to collect and use your Inputs and Outputs to develop, train and improve artificial intelligence” models unless the user opts out through account settings or the use is governed by other terms. Same editor, same keystrokes, opposite data posture by tier
  10. Cursor, by Anysphere, the tool whose own pages differ on what its privacy mode covers: the privacy overview says that with privacy mode on “Customer Data will not be used for training by Cursor” and that the company “maintains zero data retention (ZDR) agreements with all providers”, with abuse detection as the exception, while with privacy mode off it may use and store codebase data, prompts, editor actions and code snippets to train its models: cursor.com; cursor.com
  11. Anthropic, Claude Code data usage documentation, the first-party source for section 11’s two tooling findings: session transcripts are stored locally in plaintext under the user’s home directory for 30 days by default, and the feedback, bug and share commands send “a copy of your conversation history including code” to the vendor, retained 5 years, “stored in Google Cloud Storage”, with “Optionally, a GitHub issue is created in the public repository.” One limit worth carrying: the vendor’s own pages do not agree on the local retention figure, so the 30 days is what this page states rather than a settled fact
  12. Trinidad v. OpenAI, Inc., No. 4:25-cv-06328-JST (N.D. Cal. Jan. 5, 2026) (Tigar, J.), ECF No. 74, read from the RECAP copy of the order itself, 11 pages, filed 01/05/26. Unpublished, non-precedential, brought by a self-represented plaintiff and decided on the pleadings
  13. United States v. Heppner, No. 1:25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026) (Rakoff, J.), 2026 WL 436479, bench ruling 2026-02-10 and written memorandum 2026-02-17. The primary PDF was fetched, but its body is a scanned image that we did not put through optical character recognition, so only the caption, docket number, page count and filing date are verified from the primary; the holding and the body quotations are commentary-sourced with pin cites from a Paul Weiss client memo dated 2026-02-20 (not verified from a primary source): fingfx.thomsonreuters.com; paulweiss.com

Back to VPNs and TLS Protect the Pipe, Not What AI Providers Keep

Find out what your project really takes.

Book a free 30-minute consultation. You will leave with a realistic sense of size, cost and risk, whether or not we work together.