Confident Security

• The cloud platforms and model providers that already run inference could extend confidential computing and zero-retention guarantees into their own offerings, matching CONFSEC's core promise inside products enterprises already buy.
• Because OpenPCC is published as an open standard, a well-funded rival or a hyperscaler could implement it independently and stand up a compliant service competing on the same guarantees.
• The drop-in, OpenAI-compatible design and zero data retention leave little switching cost, so a customer that adopts CONFSEC could move to a cheaper private-inference option with minimal migration.
• With no named customer on the record, CONFSEC's trust claims rest on the vendor's own guarantees and a founder-stated external audit, so a security review that finds those guarantees unproven could stall adoption.
• The 2x-market-rate pricing could deter cost-sensitive buyers if a larger provider bundles comparable privacy at a smaller premium.

Confident Security treats the data a company sends to an AI model as the asset under attack. It targets the regulated enterprise in healthcare, finance, or government that wants frontier models but cannot let prompts, documents, and metadata reach a provider that might retain, read, or train on them. Its pitch is to remove that trade-off so a buyer can use AI without giving up control of the data.

The compliance page frames the gap as the guarantees a standard model API does not give. Confident Security states that customer data are guaranteed never to train models, never to be shared with a third party, and never accessible unencrypted, which are the assurances a regulated buyer needs, and which Confident Security contrasts with trusting a standard provider endpoint's policy. The problem it sells against is that using AI today means trusting a provider's policy rather than a verifiable control.

What the record does not yet carry is independent quantification of the pain. The demand case depends on the vendor's framing and on press coverage that echoes it, not on analyst sizing, survey data, or a body of incident evidence, so the problem reads as clear and credible but not yet independently measured. [s7, s1, s2]

CONFSEC is a verifiably-private inference API that runs AI models on encrypted data. A developer swaps an existing OpenAI-compatible endpoint for CONFSEC, and the service keeps prompts, outputs, and logs confidential so the operator cannot read them. The company states that prompts are never logged, retained, used for training, or sent to third parties, and that requests route through Oblivious HTTP so operators cannot link individual calls.

The design follows Apple's Private Cloud Compute and is published openly. OpenPCC ships as a framework anyone can audit and deploy, using encrypted streaming, hardware attestation, and unlinkable requests, and the homepage invites buyers to verify exactly how the code works rather than trust a promise. That auditability is the technical differentiator, because a buyer can inspect the guarantee instead of taking it on faith.

The offer spans a hosted service and self-hosting. Confident Security runs CONFSEC as a managed service on the OpenPCC standard and publishes per-token pricing at roughly twice the market rate, with public sign-up and pricing pages, and because OpenPCC is deployable on a customer's own infrastructure, a team can adopt the hosted API or run private inference in its own environment. [s8, s5, s6]

Confident Security competes on the promise of provable privacy against several kinds of rival. Confidential-computing vendors protect data in use with encryption, and data-privacy vendors strip or tokenize sensitive values before they reach a model. Confident Security differs by keeping the model's own inference private end to end rather than masking the data first.

The larger pressure comes from the platforms the company sits between. The model providers that run inference today already hold the data, and the cloud platforms that host models can add confidential computing, so the capability Confident Security sells is one the incumbents are positioned to offer themselves. Publishing OpenPCC as an open standard is the company's answer, an attempt to become the shared standard others adopt rather than a proprietary product a platform routes around.

The positioning bet is that neutrality wins. By giving the standard away and promising independent governance, the positioning reads as a bet that model providers and enterprises converge on OpenPCC, which would let Confident Security sell the managed service on top. That only holds if the standard spreads before a hyperscaler or a model provider ships an equivalent private-inference offering of its own. [s1, s5, s2]

Confident Security has a shipping product and no named customer to point to. CONFSEC is production-ready with published pricing, documentation, and an OpenAI-compatible API, but the public record names no buyer running it. TechCrunch reports the company is in talks with banks, browsers, and search engines to add CONFSEC to their infrastructure stacks, which are prospects rather than deployments.

The funding round is the strongest outside signal, and it is small. Confident Security came out of stealth in July 2025 with $4.2 million from Decibel, South Park Commons, Ex Ante, and Swyx, and by the November 2025 OpenPCC launch the seed was reported at $5 million with Halcyon and SAIF added. That backing signals investor conviction at the seed stage rather than verified commercial traction.

Distribution leans on the open standard and a self-serve product as much as direct selling. The company open-sourced OpenPCC and offers a self-serve API a developer can adopt without a sales call, a bottom-up motion that runs alongside its early enterprise conversations and has produced no paying reference customer in the public record. [s2, s3, s4]

The founder is the clearest credibility signal. Jonathan Mortensen is a two-time founder whose prior companies were acquired by BlueVoyant and Databricks, a pair of in-domain exits reported by independent press rather than asserted only on the company's own pages. That track record is the reason the team scores above the same-asset cluster.

The wider team carries relevant systems depth. The AI Journal reports backgrounds spanning Google, Apple, Databricks, Red Hat, and HashiCorp, with depth in trusted computing, secure systems, and large-scale infrastructure, which fits a product built on hardware attestation and confidential computing. The named strength is the founder's exits plus that collective pedigree.

What the record does not show is a sustained research or publication trail. The team's depth is in prior builds and employer pedigree rather than a body of peer-reviewed work or industry-standard authorship, so the credibility comes from exits and experience rather than a category-defining public reputation. [s3, s2]

Confident Security stakes its trust case on verifiable guarantees rather than certifications. The compliance page contrasts its guarantees with rivals' promises, stating that customer data are guaranteed never to train models, never to be shared with a third party, and never accessible unencrypted, and the homepage adds that the company takes financial responsibility for any breach or misuse. That liability stance is unusual and is the company's central trust argument.

The verification rests on open code and hardware attestation. Because OpenPCC is open-source and the software running inference is publicly logged and open to review, the company argues a buyer or an outside expert can confirm the guarantees rather than accept them, which is a stronger posture than a self-attested policy for a young vendor handling sensitive data.

What is absent from the public record is a named third-party attestation. The compliance page markets the ability to build SOC 2, HIPAA, PCI DSS, and GDPR compliant AI features, but a probe of it and the trust and security paths on 2026-07-03 found no SOC 2 report, ISO 27001 certificate, named auditor, or independent audit of Confident Security itself, and the founder's statement that CONFSEC was externally audited names no auditor and links no report. For a product whose entire value is trust, a security review would likely request the independent attestation the public record does not yet show. [s7, s1, s2]

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

Confident Security aims at the regulated enterprise that privacy has kept off frontier AI. The company names healthcare, finance, and government as the sectors where fear about where data goes is the barrier, and its whole pitch is to let those buyers use models without exposing prompts and documents to a provider. That buyer has both a compliance owner who needs the guarantee and an engineer who has to integrate it.

The product also courts a smaller, self-serve developer. Confident Security publishes per-token pricing and ships an OpenAI-compatible API a team can adopt by swapping an endpoint, so a developer can start without a sales call. The published price list and drop-in interface point at a bottom-up motion running alongside the enterprise pitch, which widens the buyer set but blurs which end actually pays.

A third audience is the AI vendor itself. Confident Security positions CONFSEC as an intermediary layer that model providers and hyperscalers could offer their own customers to unlock privacy-sensitive markets, so the same product is pitched to platforms whose own AI offerings could one day absorb it, a relationship the record leaves open between partner and rival. No named buyer in any of these segments appears in the public record yet.

CONFSEC runs AI inference on data the operator cannot read. It keeps prompts, outputs, and logs confidential through encryption and hardware attestation, routes requests through Oblivious HTTP so individual calls cannot be linked, and decrypts only under conditions that forbid logging, training, or human access. The company states that prompts are never logged, retained, used for training, or sent to third parties.

The differentiator is that the guarantee is auditable rather than promised. The core ships as OpenPCC, a framework inspired by Apple's Private Cloud Compute, its specification and SDKs Apache 2.0 and its reference inference server source-available under the FSL, using encrypted streaming, hardware attestation, and unlinkable requests. Because the specification and the server source are open to inspection, a buyer or an outside expert can review how the privacy works instead of trusting a policy, which is a stronger technical position than a masking layer bolted in front of a model.

The limits are the claims that rest on the vendor's own word. The founder's statements that CONFSEC is production-ready and externally audited name no auditor and link no report, and the cited sources include no third-party benchmark of the performance cost of the private path. The advantage a buyer can check is the open code and the attestation, not an independent evaluation.

Confident Security runs distribution through an open standard as much as a sales team. It published OpenPCC and framed it as an open standard any model provider or host can adopt, betting that ubiquity of the standard pulls buyers to its managed service. That is a developer-and-ecosystem motion rather than a top-down enterprise sales push, fitting a seed-stage company.

Public proof of demand is prospects, not customers. TechCrunch reports the company is in talks with banks, browsers, and search engines to add CONFSEC to their infrastructure stacks. Those are prospects that build a pipeline, but none is a paying reference the public record can name.

The funding is the strongest outside marker and it is small. The company came out of stealth in July 2025 with $4.2 million reported from Decibel, South Park Commons, Ex Ante, and Swyx, while November coverage of the OpenPCC launch reported a $5 million seed and named Halcyon and SAIF, two accounts the record does not reconcile. That backing signals conviction at the earliest stage rather than commercial traction.

Confident Security publishes its prices, which is unusual for the category and tells a buyer what it thinks it sells. The pricing page lists per-token rates for specific open models, from Mistral and Gemma at a fraction of a dollar per million tokens to larger models priced higher, updated to a dated market snapshot. Charging by the token, the same unit a buyer already uses to measure model spend, makes the cost directly comparable to a standard model API.

The company describes its pricing as twice the market rate and says so plainly, framing 2x as fair given the work required, against a rate table last updated July 9, 2025, so a buyer can see the vendor's stated premium and what it pays for. Transparent, comparable pricing lowers the friction of evaluating a young vendor, because a team can size the spend before talking to sales.

The premium is also the positioning risk. The stated 2x premium is defensible only while no larger provider offers comparable privacy closer to base cost, so the pricing depends on the privacy guarantee staying scarce rather than becoming a bundled feature.

CONFSEC is delivered as a drop-in inference endpoint. A developer swaps an existing OpenAI-compatible endpoint for CONFSEC and keeps the same interface, so the private path replaces the model API without a rewrite. Confident Security runs the inference inside hardware-isolated, attested environments and anonymizes traffic by routing it through services so the servers never see the source or content.

The company runs the standard as a managed service, and the open standard is customer-deployable. Confident Security operates CONFSEC on the OpenPCC standard as a hosted service, and because OpenPCC is publicly specified with a source-available server deployable on a customer's own infrastructure, a buyer with the capability can run private inference in its own racks. That range suits a regulated buyer that wants the data to stay inside its boundary.

Operational trust is designed to need no trust in the operator. The system retains no user data, so the delivery model removes the operator's ability to see or keep the data rather than promising restraint. What the record does not show is a published uptime commitment, support tier, or the performance overhead of the private path.

Confident Security stakes the whole product on verifiable trust rather than certification. The compliance page contrasts guarantees with rivals' promises, stating customer data are guaranteed never to train models, never to be shared, and never accessible unencrypted, and the homepage adds that the company takes financial responsibility for any breach or misuse. That liability stance is the company's central and unusual trust argument.

The verification rests on open code and hardware attestation. Because the OpenPCC specification and SDKs are Apache 2.0 and the reference inference server is source-available for review, the company argues a buyer or an outside expert can confirm the guarantees hold rather than accept a policy, a stronger posture than self-attestation for a young vendor handling sensitive data.

The gap is a named independent attestation. The compliance page markets the ability to build SOC 2, HIPAA, PCI DSS, and GDPR compliant AI features, but a probe of it and the trust and security paths on 2026-07-03 found no SOC 2 report, ISO 27001 certificate, named auditor, or independent audit of Confident Security itself, and the founder's statement that CONFSEC was externally audited names no auditor and links no report. For a product whose entire value is trust, a security review would likely require the independent attestation the public record does not yet carry.

Confident Security's platform bet is an open standard, not a closed product. OpenPCC ships as a spec and SDKs any team can implement across models and providers, plus a reference inference server and libraries for hardware attestation and unlinkable communication. The goal is a shared standard multiple model providers interoperate on rather than a single vendor's stack.

The strategy trades control for reach. By giving the core away as an open standard, Confident Security tries to make OpenPCC the default way to run private inference, which would let it sell the managed service on top and let buyers switch models or hosts without losing the guarantee. Adoption of the standard, not a proprietary lock, is the intended source of durability.

The same openness is the exposure. Because the core is openly published, a hyperscaler or a model provider can adopt OpenPCC as readily as a customer can, so the ecosystem Confident Security is trying to seed is one its largest competitors can join or absorb. No third-party marketplace or partner-built integration network beyond the open project itself appears yet.

The founder is the clearest asset. Jonathan Mortensen is a two-time founder whose prior companies were acquired by BlueVoyant and Databricks, a pair of in-domain exits reported by independent press. That record of building and selling companies in security and data is the strongest single credibility signal the company carries.

The team is built for confidential computing. The AI Journal reports backgrounds spanning Google, Apple, Databricks, Red Hat, and HashiCorp, with depth in trusted computing, secure systems, and large-scale infrastructure, which matches a product built on hardware attestation and encrypted inference. Pulse 2.0 adds institutional pedigree including Google, Apple, and Johns Hopkins.

What the record does not show is a research or standards reputation. The cited public record emphasizes exits and employer pedigree rather than published work or standards authorship in privacy or cryptography, so the credibility comes from exits and experience rather than a public track record that would itself pull adopters to the standard.