Enigmata, a cryptography startup that has operated in stealth until now, announced $6.5 million in seed funding on September 10 alongside technology it claims can train artificial intelligence models on data that never leaves its encrypted state. Blockchange Ventures led the round for a company that believes it has solved what CEO Scott Searle describes as the impossible choice enterprises face: use sensitive data to improve AI systems, or protect that data by keeping it locked away.
The timing reflects a larger collision. AI systems demand ever-larger datasets to function, yet regulations in Europe and beyond are tightening requirements around how companies handle private information. Enigmata's patent-pending Cipher system enters a crowded field where tech giants and well-funded competitors are racing to prove that computation on encrypted data can move from research curiosity to production reality.
IBM's 2026 Cost of a Data Breach report, published July 29 and covering 602 organizations, found that one in four malicious breaches now involve AI. Those incidents cost an average of $6 million, about 20 percent above the global mean. The study has sharpened corporate interest in privacy-enhancing technologies, a sector that multiple market-research firms estimate at $4 billion to $5 billion today. Projections call for compound annual growth of 20 to 27 percent through 2030, driven largely by two mathematical approaches: homomorphic encryption, which allows calculation on ciphertext, and differential privacy, which adds noise to query results.
From Theory to Cloud Infrastructure
Encrypted computation has migrated out of university cryptography departments. Google unveiled progress on its HEIR compiler for homomorphic encryption on August 14, announcing hardware-accelerator partnerships with Belfort, Niobium, Cornami, and Optalysys. The demonstrations included private recommendation engines, fraud detection, and voice-keyword recognition running entirely on encrypted inputs. Microsoft released SEAL 4.4.0 on July 25 as what it called a "critical security update," part of a steady 2026 release cadence. NVIDIA published details of its cuPQC SDK on August 29, highlighting GPU-native cryptographic operations designed for high-performance fully homomorphic encryption.
The three major cloud providers now compete on overlapping capabilities. Azure documented confidential GPU virtual machines in its NCCadsH100v5 series in June, enabling model training and inference inside hardware-enforced trusted execution environments. Google Cloud published a blog post on June 23 positioning its confidential accelerators for what it termed "verifiable trust in the AI era," particularly for agentic systems that cross organizational boundaries. AWS updated monitoring features for its Clean Rooms service on January 2 and followed with a step-by-step guide on June 8 for end-to-end encrypted machine-learning inference using Zama's Concrete-ML library on SageMaker. Both Snowflake and Google BigQuery have shipped data clean rooms with differential-privacy budgets and support for encrypted queries.
A cohort of startups is staking out specialist positions. OPAQUE Systems raised $24 million in a Series B at a $300 million post-money valuation on February 12, targeting enterprises including ServiceNow and Anthropic with what it calls "confidential AI" backed by trusted execution environments. Duality Technologies partnered with Red Hat in June to bring privacy-enhancing-technology platforms to sovereign-AI and public-sector workloads. Decentriq, which operates confidential-computing-based clean rooms, announced a partnership with Datonics on March 24 for privacy-safe audience matching. Enveil released version 6.0 of its ZeroReveal encrypted-search product on July 15, claiming throughput now limited by memory bandwidth rather than cryptographic overhead.
Regulation Tightens the Window
Europe's AI Act made Article 50 transparency obligations enforceable on August 2, and high-risk system rules take effect in December 2027. The UK Information Commissioner's Office finalized guidance on storage and access technologies in April and updated explainers on privacy-enhancing technologies, including zero-knowledge proofs, in late September. NIST published an AI Risk Management Framework concept note for critical infrastructure on April 7. Healthcare and financial-services sectors face mandates under HIPAA and sector-specific rules that academic guidance suggests are best addressed through federated learning, trusted execution environments, and synthetic data.
Cost considerations are driving adoption alongside regulatory pressure. Jeremy Kun, a staff software engineer at Google, wrote in the August 14 blog post that homomorphic encryption "fundamentally alters this trade-off by allowing computations to be performed directly on encrypted data" and that "the cost of homomorphic encryption is rapidly decreasing." Hardware acceleration underpins that trend. Optalysys raised £23 million in January for photonic computing targeting fully homomorphic encryption workloads. DARPA's DPRIVE program, involving Intel and Microsoft, continues multi-year research into dedicated silicon. Academic designs published earlier this year include FHECore, a GPU microarchitecture redesign, and processing-in-memory approaches for homomorphic operations.
Performance benchmarks show the gap narrowing, though not eliminated. A medRxiv preprint from May described FOCAL, a fully homomorphic collaborative-learning system that eliminated gradient-inversion leakage while matching state-of-the-art federated-learning accuracy. A February paper in Scientific Reports detailed Health-FedNet, which combined federated learning with Paillier homomorphic encryption and differential privacy on the MIMIC-III clinical dataset, reducing membership-inference leakage and communication overhead. An April arXiv paper demonstrated fully homomorphic encryption applied to Llama-3 inference using pipeline modifications in Concrete-ML.
Real-World Pilots Surface Trade-Offs
Actual deployments reveal both promise and limitations. Mastercard and Singapore's IMDA piloted cross-border fraud prevention using fully homomorphic encryption to check bank account numbers without revealing them, according to a Future of Privacy Forum case study most recently updated in February 2026. Google's HEIR compiler demonstrations from August included single-thread CPU latency figures and indicated that hardware-partner acceleration remains in progress. AWS's June tutorial for SageMaker walks through encrypted inference while flagging quantization and precision trade-offs inherent in current fully homomorphic encryption implementations.
Enigmata positions Cipher at the more ambitious end of the spectrum. The company said its technology "simultaneously structures and encrypts data" and that internal benchmarks show models trained on Cipher-protected data matched the accuracy of plaintext training while finishing 8 to 10 percent faster. The startup's claims are currently based on internal benchmarks and lack independent validation, as the company declined to provide independent verification at the time of the September 10 announcement. Cipher also supports targeted removal of individual records from trained models to satisfy data-deletion requests, a feature the company frames as essential for compliance with right-to-be-forgotten mandates. The longer-term vision, according to the press release, extends Cipher into infrastructure for secure data licensing and training under enforceable usage terms.
Enigmata is making Cipher available to selected enterprise design partners and uploaded a PyPI package, enigmata-cipher-tokenizer, on October 8. LinkedIn lists the company at 11 to 50 employees as of October. Searle said in a statement: "When hospitals, banks, and publishers can safely use their most sensitive assets in AI, we stop asking what AI is allowed to know and start asking what it can safely discover, prevent, and create."
The $6.5 million round attracted Blockchange for reasons that extend beyond cryptography. Matt Immerso, a general partner at Blockchange Ventures, said the firm believes "Enigmata Cipher will become a core pillar of the AI data economy, defining how sensitive data is used, licensed, and monetized in the AI era." That framing echoes claims from the broader confidential-AI cohort. Young Sohn of Walden Catalyst, an investor in OPAQUE, said in February that "enterprises will continue to struggle to bring AI into production until they have verifiable guarantees that their most sensitive data and models are protected."
What Comes Next
Encrypted AI remains workload-specific and cost-sensitive. Trusted execution environments on cloud confidential GPUs impose 4 to 8 percent throughput penalties at small batch sizes, narrowing at scale, according to a September arXiv paper. Fully homomorphic encryption for training is still emerging; most deployments focus on encrypted inference or federated learning with encrypted gradient exchange. Zama's Concrete-ML introduced encrypted fine-tuning claims in version 1.7 last year, but current documentation emphasizes inference and advises checking licensing and performance before enterprise use.
The Confidential Computing Consortium's milestones this year frame trusted execution environments as "the active foundation for modern enterprise AI," with sector use cases in finance, healthcare, advertising technology, and government moving from pilot to production. Google Cloud's June blog positioned confidential accelerators for agentic systems that cross trust boundaries. Azure and AWS documentation updated through mid-year highlights policy enforcement, attestation, and audit trails as baseline requirements for regulated workloads.
Competition will surface technical details the press releases leave unaddressed. Enigmata has not published a cryptographic whitepaper detailing ciphertext format, supported operations, security assumptions, or leakage models. The company has not disclosed whether Cipher maps to standard homomorphic-encryption schemes, integrates with compilers like HEIR, or plans support for NVIDIA's cuPQC or other accelerator SDKs. Investors and design partners will likely demand those specifics as pilots expand.
Regulatory deadlines are forcing decisions sooner than many expected. European high-risk AI systems must comply by December 2027, and August enforcement of Article 50 has already pushed content-provenance and auditable-data-use discussions forward. Aaron Fulkerson, CEO of OPAQUE, said in February that "the real barrier to enterprise AI is trust" and that "confidential AI closes the trust gap by delivering runtime proof that data stayed private." Whether cryptographic or hardware-based, that proof is shifting from optional to required for any organization deploying AI on data that carries real risk.
