Mirror Security

• No reviewed source names an organisation that has bought the product, so a partner agreement could lapse leaving no independent customer relationship behind it.
• Every performance and accuracy figure comes from Mirror, and an independent benchmark could place encrypted inference further from plaintext speed than the published band.
• Mirror's own pages state that open-weight models run unencrypted, so a buyer who reads the encryption claim as covering the whole computation could reject the product at technical review.
• Mirror carries six product lines and, on its own announcement, teams in three countries on a 2.5 million dollar pre-seed round, so it could have to drop lines before any of them reaches a reference customer.
• Intel supplies the confidential computing that the one Mirror agent deployment pattern the record documents runs inside, and a hardware or cloud platform could bundle encrypted inference into something buyers already pay for.
• Mirror states a SOC 2 Type 1 and publishes no portal a buyer can inspect, so a regulated buyer's security review could stall on evidence it cannot check for itself.

Mirror Security aims at one moment in the AI pipeline. Its pages argue that encryption at rest and in transit is settled, and that data returns to plaintext the instant a model computes on it. ThinkBusiness describes the same target, safeguarding proprietary data during model training and inference.

The buyers Mirror aims at are the ones regulation reaches. Its agent control plane ships policy packs for finance, healthcare, GDPR and HIPAA, and its testing product cites residency requirements under GDPR, HIPAA and sovereign AI mandates as the reason it needs no access to model weights or training data. No reviewed source measures how many organisations hold AI projects back over inference exposure, so the size of the problem is the company's own account.

Compliance is the frame Mirror puts around that problem. The company says it maps signed evidence to NIST AI RMF, ISO 42001, the EU AI Act, SOC 2, HIPAA, GDPR, the OWASP Top 10 for LLMs and MITRE ATLAS. That list states what Mirror aims its output at, and no reviewed source shows any of those bodies assessing the product. [s1, s3, s4, s5, s14]

Six products share one encryption layer. VectaX encrypts context, prompts, embeddings and inference output end to end, and covers encrypted vector search, encrypted keyword search and encrypted hybrid retrieval on the same key hierarchy. AgentIQ gives each agent an identity and checks every action against a natural-language policy engine and more than thirty pre-trained guardrail models. Zero inventories sanctioned and unsanctioned agents across eleven attributes each. Gateway routes model and agent calls and picks an encryption lane per route. DiscoveR runs automated adversarial testing. CodePrism encrypts code at the editor boundary so a coding assistant processes ciphertext.

The published performance numbers are Mirror's own. The VectaX page reports encrypted inference on open-weight Llama-2 models served through vLLM, with time to first token under 350 milliseconds and an accuracy loss under one point across MMLU, HellaSwag, HumanEval and GSM8K. The same page says the fuller per-configuration rows live in a technical brief, so what a reader gets is the band and not the detail behind it. No reviewed source carries a third-party benchmark of any of it.

Mirror's own pages narrow the headline claim. Mirror's published answer to why not pure homomorphic encryption is that pure FHE is too slow for production serving, and that Mirror spends its cryptographic budget on user data while keeping open-weight models in the clear. The VectaX page repeats it: the budget goes on user data, not on weights that are already open. The Intel write-up describes the same shape from another angle, naming similarity-preserving encryption for embeddings and hybrid homomorphic encryption for similarity computations. [s1, s3, s4, s5, s6, s7, s8, s10]

Mirror positions against hardware enclaves rather than against other encryption vendors. Its VectaX page sets an enclave that covers one surface, the inference enclave, against one encryption layer across every surface AI touches. The company still uses enclave hardware in its Intel work, where the Mirror software development kit is packaged inside an Intel Confidential Computing environment and attested by Intel Tiber Trust Authority.

Homomorphic encryption is a funded field in Europe and Mirror entered it small. EU-Startups grouped Mirror with several European companies working in adjacent areas of cryptography and AI security, and reported that Zama secured 49 million euro in 2025 to scale fully homomorphic encryption for confidential smart-contract execution. Mirror's own round was 2.1 million euro, against a different application of the same mathematics.

No reviewed source names a company competing with Mirror for the same buyer. What the record documents is partner relationships, so the competitive picture here comes from the shape of the category rather than from a documented contest. [s3, s10, s13]

What the record documents is partner motion rather than customer motion. Silicon Republic reported a multimillion-dollar strategic agreement with Inception AI, under which Mirror will deploy its stack across Inception's enterprise and government ecosystem, and ThinkBusiness identified Inception as part of Abu Dhabi based G42. Silicon Republic also listed earlier partnerships with Intel, MongoDB, Qdrant, SiSys AI and Accops.

The homepage presents its testimonials as direct words from operators who run AI in production with Mirror. The quotes the record captures there describe a collaboration, an integration or a programme membership instead. The NVIDIA quote calls Mirror an NVIDIA Inception member and the programme one that exists to support such companies. The Yotta quote describes integrating Mirror's capabilities into Shakti Cloud, the Accops quote describes pairing VectaX with Accops HySecure, and the Inception quote says the combination would create dependable products.

No reviewed source names an organisation that has bought anything Mirror sells. That is ordinary for a company incorporated in May 2024, and it leaves Mirror pointing at partner agreements and partner intentions for its demand evidence. [s1, s12, s14, s15]

The founding pair is publicly identifiable and the sources agree on who they are. Silicon Republic and EU-Startups both name Pankaj Thapa and Dr Aditya Narayana K as the founders, and both the company's own announcement and Silicon Republic say Mirror originated from Narayana's thesis at University College Dublin. Mirror's about page describes Thapa as spending more than two decades on enterprise and national security products, and Narayana as spending two decades across security, cryptography, privacy and machine learning in industry and academia.

Mirror's own about page carries the specifics that would lift that record above solid experience. It credits Thapa with leadership roles at Azingo, which it says Motorola acquired, Mobiliya, which it says QuEST Global acquired, and Digital14, now KATIM, which it says Edge Group acquired. It credits Narayana with 16 patents and more than 50 research publications. No reviewed independent source repeats any of those specifics.

The board adds an outside name with a matching background. Mirror announced Albert Sisto as an independent board member, describing him as a former Chief Operating Officer at RSA Data Security with more than 40 years across public companies, startups, venture capital and private equity. That announcement is Mirror's own, and no reviewed independent source covers the appointment. [s2, s9, s11, s12, s13]

Mirror states one attestation and the reviewed record offers no way to inspect it. The site says it is SOC 2 Type 1 Certified and adds GDPR and HIPAA ready. No inspectable trust portal appears on the probed surfaces, so a buyer cannot read the report or check its scope from the public record.

Mirror repeats one promise across its pages: the customer holds the keys. The company says keys stay with the customer in every deployment mode, that it deploys as software as a service, in a customer private cloud, in a sovereign region or fully air-gapped, and that nothing calls home. Its about page adds that Mirror cannot see, use or reconstruct customer data, and that customer data never becomes training material for anyone else's models.

The product's own evidence output is a claim nobody in the record has checked. Mirror says every inference and retrieval through VectaX leaves a tamper-evident signed receipt recording the request hash, the policy version, the model invoked and the encryption posture. No reviewed source reports anyone auditing those receipts. [s1, s2, s3, s16]

Add analyzed competitors to compare them side by side with Mirror Security.

Mirror aims at organisations that regulation stops from moving data. Its agent control plane ships policy packs for finance, healthcare, GDPR and HIPAA, and its testing product names residency requirements under GDPR, HIPAA and sovereign AI mandates as the reason its scans need no access to model weights or training data. The deployment options match the segment: software as a service, a customer private cloud, a sovereign region, or fully air-gapped, with the customer holding keys in every mode.

The segment the record evidences is narrower than the one the pitch names. What the cited sources document is partners in that space rather than buyers in it. Silicon Republic reported the Inception AI agreement as a route across an enterprise and government ecosystem, and ThinkBusiness placed Inception inside Abu Dhabi based G42. That is a channel toward regulated buyers rather than a regulated buyer.

Geography is part of the segmentation and part of the strain. The funding announcement puts operations in Ireland, the United States and India, and the round pays for teams in all three. For a company whose SoloCheck record puts incorporation on 1 May 2024 and its size class at micro, three geographies alongside six product lines is a wide front.

The capability that carries the company is encrypted computation, and Mirror documents it in unusual detail for its stage. VectaX encrypts context, prompts, embeddings and inference output end to end, and covers encrypted vector search, encrypted keyword search and encrypted hybrid retrieval on one key hierarchy. Its page reports encrypted inference on open-weight Llama-2 models served through vLLM, with time to first token under 350 milliseconds and accuracy loss under one point across four standard tests.

Mirror's own pages narrow what that means, and a buyer should read the narrowing before the headline. Mirror's published answer to why not pure homomorphic encryption is that pure FHE is too slow for production serving, and that Mirror spends its cryptographic budget on user data while keeping open-weight models in the clear. The VectaX page states the same trade. The Intel write-up names the mechanisms from another angle: similarity-preserving encryption for embedding storage, and hybrid homomorphic encryption for similarity computations. The encryption covers the customer's data while the open model weights run unencrypted.

The rest of the platform is conventional AI security built on top of that layer. AgentIQ pairs a natural-language policy engine with more than thirty pre-trained guardrail models and completes its runtime checks in under 25 milliseconds inline. DiscoveR runs automated adversarial testing mapped to the OWASP Top 10 for LLMs and MITRE ATLAS from outside the system under test. Zero inventories agents across eleven attributes each. The reviewed sources do not compare those with other vendors' agent controls or testing tools.

Nothing in the reviewed record validates any of it independently. Every figure is Mirror's own, the per-configuration rows live in a technical brief rather than on the page, and no third-party benchmark, audit or customer write-up appears in the cited sources. A buyer weighing the speed claim has to take the vendor's word for it or run its own test.

The go-to-market activity the record documents is partner activity. Silicon Republic reported a multimillion-dollar strategic agreement with Inception AI covering deployment across Inception's enterprise and government ecosystem, and listed earlier partnerships with Intel, MongoDB, Qdrant, SiSys AI and Accops. ThinkBusiness reported the same agreement and placed Inception inside G42.

What the partner motion has not yet produced, in any reviewed source, is a named buyer. Mirror's homepage presents its testimonials as direct words from operators who run AI in production with Mirror, and the quotes the record captures there describe collaborations, integrations and a programme membership. The NVIDIA quote calls Mirror an NVIDIA Inception member. The Accops quote describes pairing VectaX with Accops HySecure. The Inception quote is written in the conditional, saying the combination would create dependable products.

The pre-seed announcement sets the direction that spending will take. Mirror said the round would expand engineering and AI security teams in Ireland, the United States and India, accelerate product development in encrypted inference and secure fine-tuning, and drive expansion into United States enterprise markets. That is engineering teams in three countries, a United States sales push and a product roadmap funded from the same 2.5 million dollars.

None of the captured pages states a rate, a unit or what a tier contains for any of the six products, though every one carries a free-entry call to action and the homepage offers founding seats in an early-access batch. That covers the homepage, the six product pages, the about page and the three company announcements.

The absence is consistent with the motion the rest of the record shows. A partner-led enterprise sale into finance, healthcare and government does not usually publish a price, and Mirror describes deployments that range from software as a service to a fully air-gapped installation, which is a range no single list price would cover.

What a buyer cannot do from the public record is estimate cost. The captured pages carry no per-seat figure, no per-call figure and no worked example, so the encryption overhead the company benchmarks against plaintext throughput has no published price attached to it.

Mirror delivers software the customer runs, with Gateway as the hosted exception. The deployment options are software as a service, a customer private cloud, a sovereign region, or fully air-gapped, keys stay with the customer in every mode, and the company states that nothing calls home. The VectaX page frames the same choice as picking where the security boundary sits, at the gateway, at the edge, or fully on premises.

The operational claims are latency claims and they are Mirror's own. AgentIQ's runtime checks complete in under 25 milliseconds inline. VectaX reports time to first token under 350 milliseconds on open-weight Llama-2 models served through vLLM, on a rig the page describes as a single A100 for the smaller model and multi-GPU tensor parallelism for the larger one. Gateway is hosted, and the page states that VectaX runs where the data lives when it cannot leave.

Nothing in the reviewed record describes what running Mirror costs an operator. No cited source names an implementation, a support model, a service commitment or a customer's account of operating it, so a buyer sizing the operational burden has only the vendor's own latency figures to work from.

Mirror states one attestation and the record offers no way to inspect it. The site says it is SOC 2 Type 1 Certified and adds GDPR and HIPAA ready. No inspectable trust portal appears on the probed surfaces, so a buyer cannot read the report or check its scope from the public record.

The stronger trust argument Mirror makes is architectural. The customer holds the keys in every deployment mode, and the about page states that Mirror cannot see, use or reconstruct customer data and that customer data never becomes training material for anyone else's models. For a buyer whose objection is that the vendor could read the data, a key-custody design answers the objection more directly than an audit report would.

Mirror also proposes its own product as evidence. The company says every inference and retrieval through VectaX leaves a tamper-evident signed receipt, recording the request hash, the policy version applied, the model invoked and the encryption posture. It says it maps that signed evidence to NIST AI RMF, ISO 42001, the EU AI Act, SOC 2, HIPAA, GDPR, the OWASP Top 10 for LLMs and MITRE ATLAS, and that DiscoveR's scans map to the last two. Those are the company's descriptions of what its product emits. No reviewed source reports an auditor, regulator or customer acting on that output.

Mirror describes six products as one surface, and the encryption layer is what connects them. VectaX seals prompts, context and inference output. Gateway applies policy to every model call it routes and picks an encryption lane per route. AgentIQ enforces identity and policy on each agent action. Zero inventories the agent estate. DiscoveR tests the result. CodePrism encrypts code at the editor boundary so a coding assistant processes ciphertext. The company states that Gateway is hosted while VectaX runs where the data lives.

The ecosystem position in the record is a supplier position rather than a platform position. Mirror's agent deployment is packaged inside an Intel Confidential Computing environment and attested by Intel Tiber Trust Authority, so the trust root belongs to Intel. Silicon Republic lists MongoDB, Qdrant, SiSys AI and Accops as partnerships rather than as documented technical integrations, and the pairing Mirror's own captured pages describe is VectaX with Accops HySecure, though a Mirror article headline in the record also names a MongoDB partnership. Inception AI is the named route to an enterprise and government ecosystem.

That arrangement cuts both ways and the record does not settle which way it cuts harder. Sitting under other vendors' AI stacks is how a small company reaches regulated buyers it could not sell to directly. It also means Intel, whose confidential computing hosts the one Mirror agent deployment pattern the record describes, sells an adjacent control of its own, and no cited source describes a commitment that would keep Mirror in that position if Intel chose to change course.

The founders are identifiable and the sources agree on the pair. Silicon Republic and EU-Startups both name Pankaj Thapa and Dr Aditya Narayana K, and both the company's own announcement and Silicon Republic say Mirror originated from Narayana's thesis at University College Dublin. Mirror's own page says Thapa leads strategy, partnerships and sales after more than two decades on enterprise and national security products, and that Narayana leads research and product direction after two decades across security, cryptography, privacy and machine learning.

The claims that would make this an unusually strong founding team come only from Mirror. Its about page credits Thapa with leadership roles at Azingo, which it says Motorola acquired, at Mobiliya, which it says QuEST Global acquired, and at Digital14, now KATIM, which it says Edge Group acquired. It credits Narayana with 16 patents and more than 50 research publications, and describes the wider team as people who have shipped enterprise and nation-state security software for two decades. No reviewed independent source repeats any of those specifics.

The board adds an outside name with a matching background. Mirror announced Albert Sisto as an independent board member, describing him as a former Chief Operating Officer at RSA Data Security with more than 40 years across public companies, startups, venture capital and private equity. The announcement is Mirror's own, and no reviewed independent source covers it.

Six product lines, three geographies and a 2.5 million dollar round is a wide brief for a company SoloCheck's record classifies as a micro company, and the record does not yet show that brief being met.