Law No. 6698 (amended by Law No. 7499)
Under Law No. 6698 and its secondary legislation, these policies keep personal data and special-category personal data safe across AI interactions. Coverage reflects the 2024 Law No. 7499 amendments (special-category grounds, cross-border transfer, the administrative-fine regime) and the Board's AI guidance: the AI Recommendations (2025), the Generative AI Guide (Nov 2025), the workplace generative-AI announcement (Mar 2026) and the Agentic AI document (2026).
Coverage focus
Detects the 11-digit Turkish national ID (TCKN) pattern in prompts and model outputs using algorithmic validation. A checksum verification avoids collisions with random 11-digit numbers; single detections are masked, bulk detections are blocked.
Uses Turkish named-entity recognition (NER) to detect real persons' full names in a context-aware way. Public figures are distinguished from data-subject individuals; customer and employee names are masked before being sent to external models.
Detects and masks phone numbers in +90, 05xx and international formats. Corporate switchboard numbers can be whitelisted.
Detects and masks personal email addresses. Corporate domains are managed with an exception list; personal domains such as gmail and hotmail are masked by default.
Detects residential addresses, neighborhood/street/door-number combinations and sensitive location data belonging to real persons. Corporate locations such as facility addresses are excluded.
Detects cases where individually anonymous-looking data (e.g. birth date + province + occupation) combine to make a person identifiable. Raises a warning when the combination threshold is exceeded, applying the re-identification-risk guidance from the KVKK AI Recommendations at runtime.
Detects health data such as diagnoses, medication, reports, hospital records and disability status. As special-category data the default action is block. After Law No. 7499, explicit consent is no longer the sole legal ground under Art. 6/3; exceptions for legal obligations in employment, occupational health & safety and social security open only through a justify workflow in authorized HR/OHS projects, and each exception logs the Art. 6/3 clause relied upon.
Detects and blocks data on individuals' religion, sect or other beliefs. General religious and cultural topics are allowed; belief data attributed to a specific person is blocked.
Detects and blocks political views, party membership or philosophical belief data attributed to a specific person.
Blocks processing of association, foundation or trade-union membership data of employees or third parties in AI systems. Anonymized use for collective-labor-law advisory scenarios is allowed.
Detects and blocks references to fingerprints, face-recognition templates, retina, voice biometrics and genetic data. Pasting biometric configurations from facility security systems into prompts is also caught.
Blocks processing of individuals' criminal records, convictions and security-measure information. Entering criminal-record checks into AI tools during recruitment is a typical violation scenario.
Detects and blocks race and ethnic-origin data attributed to specific individuals.
Prevents special-category data on appearance/dress and sex life (listed in KVKK Art. 6) from appearing in AI interactions.
Detects TR IBAN, credit-card (Luhn-validated), account-number and tax-ID patterns. Card numbers are blocked in all cases; IBANs may pass masked in corporate payment scenarios.
Detects personnel data such as registry numbers, social-security numbers, payroll, salary, performance notes and disciplinary records. Entering this data into AI tools outside authorized HR projects is blocked; processing relying on the Art. 6/3 employment exception is logged with a justification.
Blocks when records of more than one person (customer lists, Excel pastes, CSV dumps) are detected in a single prompt or output. The threshold is configurable; the default blocks above 3 records.
Blocks — or passes only in masked form — requests containing personal data that would be routed to model endpoints hosted abroad. Under the post-7499 Art. 9 regime, any personal data sent to an external model endpoint counts as a cross-border transfer and must rely on a transfer mechanism (adequacy decision, standard contract, binding corporate rules or an undertaking); endpoints without a defined mechanism are locked by default (see KVKK-23).
Detects use of personal data outside the processing purpose defined for the project — for example processing personal data for marketing segmentation inside a customer-support project. Applies the KVKK AI Recommendations' principle that if a result can be reached without processing personal data, personal data must not be processed.
Detects Art. 11 applications (erasure, rectification, information requests) in chatbot interactions and routes them to the official application channel. The AI is prevented from answering these requests on its own; person-level erasure is supported via the audit ledger, and an incident record tracks the 30-day response window under Art. 13.
Detects and blocks content indicating that personal data of persons under 18 is being processed. In all scenarios where parental or guardian consent cannot be verified, the default action is block.
Records every AI interaction containing personal data in the hash-chained audit ledger and reports which data category was processed in which project in a format comparable with the VERBIS declaration. Exported as an evidence pack during a KVKK inspection.
Each external model endpoint and third-party service is mapped in Control Center to a defined Art. 9 transfer mechanism (adequacy decision, the standard contract published by Board decision 2024/959, approved binding corporate rules, or a Board-authorized undertaking). Personal-data traffic to endpoints whose mechanism is missing, expired, or not registered as notified to the Board within 5 business days is blocked; the notification date is kept as mandatory metadata.
When it detects personal-data breaches through AI interactions (prompt-injection-driven disclosure, bulk leak, unauthorized-access pattern), it opens a high-priority incident record and compiles the fields required by the KVKK breach-notification form. The 72-hour timer starts from the moment the breach is learned, an instant notification is sent to the CISO, and the moment of detection is stamped provably in the hash-chained ledger.
Allows personal-data traffic only to model endpoints that contractually commit that inputs will not be used in training (zero data retention / no-training). Cuts off — at the source — the risk highlighted in the Board's Generative AI Guide of user inputs leaking into model training; personal data can reach endpoints without such a commitment only after full redaction.
Inspects employees' entry of personal and corporate data into public generative-AI tools such as ChatGPT, Gemini and Claude at the browser layer through the UMAI Extension. Use of individual or free accounts outside the approved corporate subscription is detected, and inputs containing personal data are masked or blocked — directly addressing the Board's 05.03.2026 workplace generative-AI announcement.
Inspects personal-data flow in autonomous agent systems' tool calls, memory writes and agent-to-agent messaging. The scope of personal data an agent can access is defined per project; out-of-scope access attempts, writing personal data to agent memory and sharing personal data between agents are logged as separate events, covering the risks flagged in the Board's 2026 Agentic AI document.
Assigns project-based retention periods to personal data in AI interaction logs; records whose period expires are deleted, destroyed or anonymized in a periodic destruction cycle, with the destruction written to the audit ledger as evidence. Person-level erasure requests (Art. 7) are applied on the ledger so data whose purpose has ended does not persist through logs.
Detects outputs based solely on automated analysis that produce adverse consequences for a person. Such outputs require mandatory human review, and the decision rationale and input scope are recorded so the data subject can exercise their objection right under Art. 11/1-g. HR pre-screening, supplier evaluation and customer scoring projects bind to this policy by default.
Continuously measures the data categories, processing volume, special-category-data contact rate and high-risk indicators of every AI project that processes personal data, producing structured input for the privacy impact assessment the KVKK AI Recommendations prescribe for high-risk processing. A re-assessment alarm triggers when a project's risk profile changes, and it feeds the KVKK component of the Governance Health Score.