DataPacts

Security During an AI Data Sale: Access, Encryption, Breach Risk

Checked

At a glance

Security during an AI data sale is the set of access limits, encryption duties and breach rules that protect company data from the first sample to the buyer's deletion of the raw copy.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Four points carry the exposure: access, transfer and storage, written promises, and breach. Buyer copies outlive the deal unless the contract says otherwise.

In AI training data licensing, the written terms are the only security a seller can enforce: program-page statements are claims. The published master contract reviewed here, micro1's Data License Master Terms (last updated 2026-11-12, read 2026-11-16), sets encryption, a two-business-day incident notice and flow-down duties.

Fact-check note: every term, date and figure on this page carries its source and check date, and the method is described under how we source, date and check every figure.

What Security Covers When Company Data Is Licensed

In a data license, security means controlling who can reach the raw records at each step, how the data is protected while it moves and while it is stored, and what each side must do if it is exposed. Four areas recur:

  • access to the raw records;
  • encryption in transit and at rest;
  • incident notice and cooperation;
  • copies held by affiliates, customers and in backups.

Where Company Data Is Exposed, Stage by Stage

Company data is exposed at seven points of a sale, growing from answers about the data at intake to full raw records once a program extracts or receives them. The table is as of 2026-11-16 (program values checked 2026-11-09, micro1's terms read 2026-11-16).

Stage What may leave the company Who can hold a copy Control on record Source
Intake and estimate answers about size, systems and history the program or intermediary Troveo states "No OAuth access or install required" troveo.ai/business-data
Sample a redacted subset the buyer's evaluators Patronus AI states samples are used only for evaluation and deleted with written confirmation if there is no deal data.patronus.ai
Diligence documents about the data, not the data the buyer's reviewers Defined.ai lists cybersecurity among its compliance checks defined.ai/partnerships
Extraction and de-identification raw records, if the program extracts program, intermediary or an independent agent read-only connection (Magma, Pangea.ai); in-environment, on-prem or VPC options (Scale AI); an independent Deidentification Agent in Spirit magmahq.ai; pangea.ai; scale.com; courtlistener.com
Transfer the licensed dataset in transit delivery by an agreed method with export cost on the seller (micro1 s.1.1); a secure-upload exhibit in Spirit micro1.ai master terms; courtlistener.com
Use at the buyer the dataset and copies at affiliates or customers buyer, affiliates, customers, service providers encryption in transit and at rest, need-to-know access, flow-down (micro1 s.2.4, 2.6) micro1.ai master terms
After the term backups, trained models the buyer raw data deleted within 120 days, backups excluded, models kept (micro1 s.4.4); Labelbox states trained models "cannot be unlearned" micro1.ai master terms; prism.labelbox.com/faq

The early stages can run on redacted subsets, as covered under proving data quality without sharing the data.

Who Touches the Raw Data: Three Access Models

Programs reach company data in one of three ways: a read-only connection, processing inside the company's own environment, or an export the company uploads itself.

Read-Only Connections to Company Systems

A read-only connection lets a program extract records without changing them, so the company limits what is copied by scope rather than by what it sends. Magma, a marketplace, states a read-only encrypted connection limited to an approved scope.

Pangea.ai, also a marketplace, states a company connects GitHub, GitLab or Bitbucket read-only. Centific states the company connects in-scope systems read-only and that "You license your data. You never hand it over." (all checked 2026-11-09).

Mercor states the steps Scope, Connect and Process and that "No AI lab buyer or third-party vendor ever sees your raw data" (checked 2026-11-09). Connector types, API terms and revocation sit under buyer connectors versus native exports.

Processing Inside the Company's Environment

Some programs de-identify the data before it leaves the company, so the raw records stay inside the company's own systems or private cloud. Scale AI states it anonymizes inside the company's environment before extraction, with per-package approval gates and on-prem or VPC options.

Replay states each engagement runs in an isolated environment and only de-identified data is licensed (both checked 2026-11-09). These are program statements, not contract terms.

Seller-Side Export and Upload

When the company exports the data itself, no outside party touches its systems, and the security question moves to how the export travels and who can open it. Troveo, a marketplace that licenses on the seller's behalf, states a package is anonymized with a manifest (checked 2026-11-09).

In micro1's terms, delivery runs through an agreed method and the licensor bears export and delivery costs (s.1.1, read 2026-11-16). Formats and acceptance are covered under delivering licensed data to an AI buyer.

Security Terms in Published Buyer Contracts

Security promises bind only where they sit in the signed terms, and micro1's Data License Master Terms are the published buyer master terms on record that set them clause by clause. The version posted on an order form's effective date governs it.

What the Buyer Promises About Encryption and Access

micro1 commits to commercially reasonable administrative, technical and physical safeguards (read 2026-11-16). Four clauses carry the security duties:

  • s.2.4: encryption of the data in transit and at rest, need-to-know access controls and written information security policies;
  • s.2.5: notice within 2 business days after becoming aware of material unauthorized access or disclosure, with cooperation at the licensee's cost;
  • s.2.6: before data goes to a permitted affiliate or customer, the licensee ensures it is bound to substantially equivalent security obligations;
  • s.1.4(a): no sublicense to a customer for AI training without reasonable measures to redact, anonymize or pseudonymize PII.

The master terms contain no audit or exclusivity right (s.4.2, checked 2026-11-09). Mercor's data page states a SOC 2 Type II report and AWS KMS keys (checked 2026-11-09) but no license terms, so these are claims. The clause-level view is in the security clauses of an AI data licensing agreement.

What the Seller Promises About the Data It Hands Over

Security duties run both ways, because the seller warrants that the data carries no malicious code and no hidden way to reach or disable the buyer's copy. In micro1's terms (read 2026-11-16, indemnity checked 2026-11-09):

  • s.5.7(d): the data contains no viruses, malware or other harmful material;
  • s.5.9: no device or mechanism lets the licensor or a third party remotely access or disable the licensee's use;
  • s.8.1: the licensor indemnifies the licensee for IP, privacy, publicity and data protection claims arising from the data.

Breach Risk After the Data Leaves

Once a copy sits with a buyer, the seller's protection against a breach is the notice it receives, the remedies it keeps and the duties the law puts on whoever holds personal data.

Notice Clocks in Contracts and Law

A buyer's breach reaches the seller only as fast as the contract requires, and the clocks on record range from 24 hours in a non-binding EU model term to 30 days for a US regulator (table as of 2026-11-16, ordered by deadline).

Duty Who must notify whom Deadline Status Source
EU model contractual term, Annex V clause 4.4.4 each party to the other (Annex V model terms for voluntary data sharing) 24 hours of discovery non-binding Recommendation (EU) 2026/2394, adopted 2026-10-29, OJ 2026-11-06; applies only if written into the contract eur-lex.europa.eu
micro1 Master Terms s.2.5 licensee to licensor, with cooperation at the licensee's cost 2 business days after becoming aware of material unauthorized access or disclosure contract term, read 2026-11-16 micro1.ai master terms
GDPR Art. 33(2) processor to controller without undue delay law, applies from 2018-05-25 publications.europa.eu
GDPR Art. 33(1) controller to supervisory authority without undue delay and, where feasible, within 72 hours, unless unlikely to result in a risk law publications.europa.eu
GDPR Art. 34(1) controller to data subjects without undue delay, where likely high risk law, read 2026-11-16 publications.europa.eu
GLBA Safeguards Rule, 16 CFR 314.4(j) covered financial institution to the FTC no later than 30 days after discovery, events of at least 500 consumers federal rule, as of 2026-11-05 ecfr.gov

The GDPR and GLBA clocks run to authorities and people; only the contract clock runs to the seller.

Who Pays and What Ends if a Breach Happens

A breach claim against a buyer runs into the contract's liability cap unless one of its carve-outs applies. micro1's cap is 3 times the fees paid or payable under the order form for each party, except fraud, gross negligence and willful misconduct (s.7.2, read 2026-11-16).

Cap mechanics are in liability caps in AI data licenses. Under GDPR Art. 82(4), each responsible controller or processor is liable for the entire damage, with recourse under Art. 82(5).

A breach can also end the license: micro1's terms allow termination after a 10-day cure period (s.4.1), and Labelbox Prism states deletion within 30 days on triggers including material breach (both checked 2026-11-09).

What a Buyer's Security Record Tells a Seller

A buyer's security record is public only where an incident reached a regulator or a court, and one data buyer, Mercor, has such a record.

Mercor reported unauthorized access through compromised LiteLLM versions (2026-03-24 and 2026-03-30) to the California Attorney General on 2026-06-25. Mercor states a "very limited subset" of "nearly five million" experts had sensitive information affected. That is experts' information, and no incident exposing licensed company data is on record in the sources cited.

The lead case is In re Mercor.io Corporation Data Breach Litigation, N.D. Cal. No. 3:26-cv-02821-WHO, with the parties' stipulation of 2026-07-21 to stay it pending mediation (Dkt. 44), with no ruling on the merits as of 2026-11-09. The dated list is the register of data buyer incidents.

Where the Law Sets a Security Floor

Where licensed data holds personal data of people in the EU, GDPR Art. 32 requires security appropriate to the risk from controller and processor, naming pseudonymization and encryption. Art. 32(1)(a) to (d) lists pseudonymisation and encryption, confidentiality, integrity, availability and resilience, timely restoration, and regular testing (read 2026-11-16).

Art. 28(3)(c) requires the processor contract to oblige the processor to take all measures required under Art. 32.

A proposed HIPAA Security Rule update (90 FR 898, 2025-01-06, RIN 0945-AA22) would raise requirements for health data holders. It is proposed, not in force; the Unified Agenda lists final action for 07/2027. The sale rules are under HIPAA and AI training data.

Under 28 CFR 202.302, in force since 2025-04-08, data brokerage with a foreign person requires a contractual bar on onward transfer to countries of concern and reporting of violations within 14 days. Export-control rules treat end-to-end encrypted sending or storing as not an export (22 CFR 120.54(a)(5), 15 CFR 734.18(a)(5)).

Security Questions to Settle Before Any Data Leaves

Eight questions turn security from a program's claim into a term the company can enforce, each answered in the order form or master terms before data moves.

  1. Which access model applies?
  2. Where do the raw records sit during processing?
  3. Does the contract name encryption in transit and at rest, and need-to-know access?
  4. What is the notice period, and how is a security incident defined?
  5. Are affiliates and customers bound to equivalent duties (flow-down)?
  6. Is sublicensing for AI training tied to PII measures?
  7. When are raw data and backups deleted, and is it confirmed in writing?
  8. Can the company audit the buyer or obtain written confirmation?

The last question belongs to audit rights in AI data licenses, a page not yet published, because micro1's master terms leave audit to the order form.

Tighter Controls, Standard Terms or No Sale: Benefits, Risks, Downsides and Alternatives

Tighter security terms lower exposure but not to zero, and declining to license is the only option that keeps every copy inside the company.

Benefits

  • Written encryption, notice and flow-down duties turn claims into enforceable terms.
  • An approved scope or in-environment model keeps raw records with the company.
  • A defined notice period lets the company act on a buyer's incident.

Risks

  • A cap of 3 times fees can leave a large breach loss unrecovered.
  • Copies at affiliates, customers and in backups sit outside the company's control.
  • Labelbox states trained models cannot be unlearned, so deletion does not undo use.

Downsides

  • Negotiating security terms takes time and may narrow the buyers willing to proceed.
  • In-environment processing and de-identification place work and cost on the company.
  • A page statement does not bind a buyer unless written into the terms.

Alternatives

  • License a de-identified subset.
  • Choose an in-environment model so raw records never leave.
  • Delay the sale for a legal review of the security terms.
  • Decline and keep the data, as covered in alternatives to selling data to AI labs.

Where Security Sits in the Process to Sell or License Company Data

Security runs through every step of the sale, so each step below carries its own access limit. The guide to sell or license company data to AI companies covers the full process:

  1. NDA before sharing data
  2. Data room for an AI data licensing deal
  3. Due diligence in an AI data deal
  4. PII removal and de-identification
  5. Delivery to the buyer
  6. After the sale

Security Arrangements in a Court-Supervised Data Sale

In the Spirit Airlines bankruptcy, the security of the data being sold was part of the court record, not only of the contract. Passenger databases are excluded, Tonic.ai is hired as Deidentification Agent (Dkt. 1581, 1684), and an alternate sale agreement provides delivery by secure upload (Exhibit C).

The Consumer Privacy Ombudsman's Supplemental Report (Dkt. 1684, filed 2026-10-05) asks for further research into Mercor's security incident. The successful and alternate bids (Dkt. 1463, 2026-08-14) are not approved as of 2026-11-09, with a hearing on 2026-11-17 (Dkt. 1779). The rules are under consumer privacy ombudsman in US bankruptcy data sales.

This page is information, not legal advice.

Programs and Terms Cited on This Page

The programs and documents cited above, in alphabetical order (the Centific profile is not yet published):

Frequently Asked Questions

Is it safe to license company data to AI companies?

It carries real exposure, because the buyer keeps a copy, may pass it to affiliates and customers under flow-down terms and keeps backups after deletion, so safety depends on the access model and the written terms.

The decision as a whole is covered under whether to sell or license your data.

Do AI data buyers encrypt licensed data?

Only some say so in their terms: micro1's master terms require encryption in transit and at rest, while other programs make page statements that are not terms.

Does a buyer have to tell the seller about a breach?

Only if the contract says so; micro1's terms set two business days, while GDPR's clocks run to authorities and individuals, not to the licensor.

Can an AI data buyer connect directly to company systems?

Some programs ask for a read-only connection limited to an approved scope, others process inside the company's environment, and some need no access because the company exports the data.