Enigmata says its encrypted training beats plaintext by 8 to 10 percent

Privacy-preserving computation has historically cost performance, most of all in systems built on fully homomorphic encryption ^[14]. Enigmata's figure points the other way. Measured against training on raw data, the protected run is reported as finishing 8 to 10 percent sooner at matching accuracy ^[12], so the claimed overhead of protection is negative, somewhere between -8 and -10 percent ^[21]. Enigmata's own testing is the only source for it ^[12]. At seed stage that is ordinary; it is still one run by one interested party.

For that percentage to transfer to another shop, the two runs need the same model architecture and hyperparameters, the same hardware, the same step count to whatever counts as converged, and the cost of applying the transform paid outside the measured window. If the protected representation changes input dimensionality or field cardinality, the two runs are two different workloads and the percentage says nothing about protection. Enigmata has not published the workloads, model configurations, datasets or the independent security analysis that would let anyone check ^[13].

Compatibility is the pitch. Cipher turns fields, records, documents and feature sets into protected representations that Enigmata says standard machine-learning, analytics and search systems can still use ^[5], on existing enterprise hardware ^[8], with support described across Databricks, SageMaker, Snowflake, Vertex AI, Azure Machine Learning, on-premises GPUs, private clouds and controlled data centers ^[9]. An unmodified pipeline consumes whatever structure the transform left intact. That structure is also what anyone holding the protected data has. Enigmata's public materials do not give enough architectural detail to compare its security assumptions with fully homomorphic encryption, secure multiparty computation or trusted execution environments ^[10].

The neighbours are legible by comparison. Duality Technologies markets encrypted analytics and machine-learning workflows using fully homomorphic encryption and secure multiparty computation ^[15]. Opaque keeps data encrypted in memory and uses attestation to verify workloads ^[16]. Enigmata describes Cipher as a separate cryptographic transform designed to preserve data utility on standard infrastructure ^[20].

Searle put the thesis plainly in the company's announcement: "The world's most valuable data no longer has to stay locked away to stay protected." ^[11] He previously founded LockStream, a digital-rights-management and encryption company, and patent records identify him as an inventor on LockStream encryption and content-accountability technology, including a patent involving encryption with personally valuable user keys ^[18].

Conventional encryption protects data at rest and in transit, then needs an authorized system to decrypt before meaningful computation starts, and that plaintext stage is the exposure Enigmata wants out of AI workflows ^[17]. In a Cipher deployment the stage moves to the reveal. Policy controls decide when an approved user or system can reveal source fields, and each reveal is recorded in an audit trail ^[7]. Intended workloads include model training, inference, semantic search, retrieval-augmented generation, fraud analysis and third-party data collaboration ^[6]. For records that should never be restored to their original form, Enigmata markets a second product, Anonymizer, a one-way transformation ^[19].