Blind spots created by raw-data limits
Insurers, financial institutions, and platforms often decide with only their own signals, so duplicate claims, identity mismatch, and anomalous patterns are detected late.
B · B2B
Private risk checks across institutions
Insurance, finance, and platform risk decisions improve when signals can be checked across institutions. PET computes the needed risk result without moving raw customer data.
B · B2B
Customers buy the risk decision, not the raw data
This product track privately cross-checks claims, identity mismatch, FDS, and anomalous transaction signals that were previously blocked by raw-data sharing limits.
Private cross-check result
Claims, identity, account, and transaction patterns remain with each institution while the customer receives only the risk score and validation result needed for action.
Use-case design to subscription or per-check pricing
The commercial path starts with use-case design and validation, then can move into subscription or per-check pricing once repeat usage is proven.
B · B2B · Agent Security
FHE16 Agent Gate: controlling what an AI agent is allowed to do
AI agents now execute payments, refunds, data exports, and risk actions directly. Agent Gate sits between the agent and its tools as an execution-security layer that enforces pre-approved rules and returns only the needed decision without exposing sensitive inputs.
When data becomes an instruction
When an agent reads external documents, email, or web content, embedded malicious instructions can chain allowed tool calls into data exfiltration or wrong actions. In the NIST CAISI 2026 red-team, all 13 targeted frontier models had at least one successful attack.
A three-layer execution boundary
① A deterministic policy checks allowed tools, identity, amount, recipient, purpose, and evidence; ② protected amounts, thresholds, and risk scores are evaluated on ciphertext with FHE16; ③ a threshold-approval step keeps any single party from decrypting alone. The output is ALLOW / REVIEW / DENY with a reason code.
High-risk actions in finance
The first targets are payment, refund, PII export, and limit/risk actions, where an error costs more than a wrong answer. In Korea, 169 generative-AI innovative financial services have been designated and 45 launched — read as an adoption signal, not a customer count.
What works now vs. the pilot scope
A deterministic policy gate works today (9/9 baseline tests). The on-ciphertext decision and threshold release, plus a Naver Cloud CLOVA Studio Skill connection with domestic-region performance and cost measurement, are the pilot scope. This is an integration in progress that links a policy baseline with cryptographic research assets, not a finished product.