Openlayer

• Databricks, which already governs AI traffic through Unity AI Gateway, could build equivalent guardrails natively and drop Openlayer from the partner path.
• The clouds and model providers that host the models could offer routing, spend caps, and content guardrails in their own endpoints, removing the reason to add a separate gateway in front of them.
• Openlayer addresses security and compliance teams with the gateway while its evaluation product addresses engineers, and it could fail to build a security sales motion alongside the developer one it already has.
• Output guardrails do not run on streamed responses, so buyers whose applications stream by default get input-side control and no response-side control.
• A prompt-injection guardrail that always blocks on a confidence threshold can produce false positives, and teams may raise the threshold until the control stops firing.

Openlayer aims the gateway at companies whose teams have already put copilots, agents, and models into production faster than their controls could follow. Its homepage states the gap plainly, saying most companies cannot list the AI systems they run, few can test them, and almost none can prove to a regulator what any of them did.

The buyer it names for this line is the security and compliance function. Openlayer's page for that audience describes keeping a current inventory of AI systems and enforcing policy on live interactions, alongside the framework mapping its governance product handles. That is a budget line a security leader recognizes.

The quantified pain on those pages comes from outside research the vendor republishes rather than from measurement of its own buyers, so the problem statement is specific about who hurts and general about how much. [s21, s9, s3]

The gateway is a single endpoint that applications point their base URL at, keeping the OpenAI- or Anthropic-compatible client they already use. Every request is checked in a documented order, cheapest check first: model access, then usage limits, then content guardrails. Anything refused is stopped there, before it reaches a provider.

The guardrails are the security half. A personal-data guardrail takes two entity lists, one whose matches reject the request outright and one whose matched spans are replaced with placeholders before the request is forwarded, with a confidence threshold that tunes how sure the detector must be. Prompt injection takes a threshold and always blocks, and it scores only the user's turns, which the vendor explains by saying system prompts routinely contain phrasing a classifier reads as an injection while assistant turns and tool results are not user-controlled. Content safety scores six categories, and it has no redaction path, so the outcome there is block or allow.

Openlayer documents the limits alongside the features. Output guardrails do not run on streamed responses, since a streamed token reaches the caller before a complete response exists to inspect. The violation log records metadata rather than the text that triggered it. [s2, s1, s11, s12]

The scored line sits in a control point the platforms want as well. Databricks describes Unity AI Gateway as its governance solution for enterprise AI and opened it in June 2026 to outside AI security providers, naming Openlayer among them beside CrowdStrike, Palo Alto Networks, and Zscaler. The reviewed sources do not establish what competing gateway vendors ship.

Openlayer's own answer is the join with the rest of its platform: the guardrails run the same test definitions as its evaluation and monitoring products, so what a team tested before shipping is what the gateway enforces afterward. The reviewed sources do not establish how hard that join would be for a rival gateway to reproduce.

The Databricks relationship shows the position from the other side: Openlayer's guardrails sit inside a platform control point as well as behind its own. [s16, s4, s3]

The company's commercial evidence is real and sits mostly next to the gateway line rather than under it. SiliconANGLE reported that Amdocs and Telefónica use Openlayer's tools to evaluate their AI models, and Mobile World Live reported in March 2026 that Telefonica Tech would fold Openlayer into its AI and data services across Europe and Latin America, contributing consulting, systems integration, and managed operations.

The published customer stories are the same shape. The Jericho Security account describes a monitoring mode that analyses large volumes of outputs and a development mode for benchmarking prompt sets, which is the evaluation product at work. No reviewed source names a customer of the gateway. What does belong to the gateway line is the Databricks relationship, where Openlayer supplies observability, evaluation, and runtime guardrails into governed AI workflows.

Distribution is broadening at the company level rather than for the gateway line specifically. Openlayer's changelog records availability on the AWS, Azure, and Google Cloud marketplaces, and its published plans run from a free Basic trial capped at 20,000 inference logs a month to a negotiated enterprise plan adding role-based access control, on-premises deployment, and compliance reports. The published plans name neither the gateway nor the guardrails. [s14, s15, s10, s16, s8, s7]

Gabriel Bayomi, Rishab Ramanathan, and Vikas Nair founded Openlayer in 2021 and hold the chief executive, technology, and product roles. SiliconANGLE reports independently that the three worked together on Apple's machine-learning efforts, and the company says that work meant building models for Siri and Vision Pro while rebuilding the same evaluation infrastructure team by team.

The company came through Y Combinator's Summer 2021 batch, which lists a team of 24. Race Capital led its Series A, with KPN, Y Combinator, NXTP, Mindset, Quiet Capital, and Telefónica also taking part.

That background is machine-learning evaluation rather than security engineering. The reviewed record shows no prior security product, exit, or research record behind the guardrails the gateway now enforces. [s5, s14, s20]

Openlayer publishes a trust center at trust.openlayer.com and a security page stating SOC 2 Type II, GDPR, and HIPAA, with penetration tests run by independent security firms and continuous posture monitoring through Vanta. A random control subdomain of the same domain did not resolve, so the trust surface is a real host rather than a wildcard.

Enterprise deployment options are documented: air-gapped installation inside a customer's own network with no external connections, data residency in the European Union or the United States, single sign-on through SAML and OIDC, directory sync, and role-based access control on the paid plan.

These are the controls a regulated buyer expects to find rather than ones a competitor would struggle to obtain. They clear procurement and they do not block a replacement. [s6, s13, s7]

Add analyzed competitors to compare them side by side with Openlayer.

Openlayer points the gateway at security and compliance teams inside companies that already run AI in production across business units. Its page for that audience frames the job as keeping a current inventory of AI systems, turning EU AI Act, ISO 42001, NIST AI RMF, and OSFI E-23 requirements into controls with owners, and enforcing policy on live interactions.

Openlayer addresses engineers as well. Its evaluation product is presented for testing language models, agents, retrieval systems, and traditional machine learning, and the named accounts in the reviewed record sit with that product rather than with the gateway. The reviewed sources do not settle whether the same team buys both.

The segment it addresses is regulated enough to need the deployment options Openlayer publishes, including installation inside a customer's own network with no external connections and data residency in the European Union or the United States. The reviewed sources do not attribute a named regulated buyer to the gateway line itself.

The mechanism is a proxy that applies policy to every request as it passes through. Applications keep their OpenAI- or Anthropic-compatible client and point its base URL at Openlayer, which verifies the key and then checks model access, usage limits, and content guardrails in that order before forwarding anything.

Machine learning does the content work. A personal-data guardrail detects entity types from universal ones such as email addresses and payment-card numbers through to country-specific identifiers, rejecting a request that contains a blocked entity and replacing matched spans of a redacted entity with placeholders before forwarding. A prompt-injection detector scores the user's turns against a confidence threshold and always blocks, deliberately skipping system prompts, which the vendor says routinely contain phrasing a classifier reads as an injection, along with earlier assistant turns and tool results, which it describes as not user-controlled. Content safety scores six categories and blocks above a per-category threshold, with no redaction path available.

The advantage Openlayer claims for itself is continuity rather than detection quality. The guardrails run the same test definitions as its evaluation and monitoring products, so a policy a team tested before shipping is the policy enforced afterward, and each enforcement decision lands in the same trace the monitoring product already shows.

Adoption is meant to cost a configuration change rather than a project. A team swaps a base URL and a key, and the policies are set once in an admin portal with no redeploy of the applications sending traffic. That is a low-friction entry into an account where Openlayer may already be running.

The visible motion is partner-led as much as direct, and it is company-level rather than specific to the gateway line. Mobile World Live reported in March 2026 that Telefonica Tech would fold Openlayer into its AI and data services across Europe and Latin America, supplying consulting, systems integration, managed operations, and local support. Databricks named Openlayer among the AI security providers integrating with Unity AI Gateway. Openlayer also lists on the AWS, Azure, and Google Cloud marketplaces.

What the record does not show is a customer of the gateway. The named accounts, Amdocs and Telefónica in SiliconANGLE's reporting and Jericho Security in Openlayer's own case study, all describe the evaluation and monitoring products.

Openlayer publishes no price for the gateway or for the guardrails. Its pricing page describes evaluation, testing, observability, and tracing features and names neither, listing a free Basic trial capped at 20,000 inference logs a month with one member, five projects, and 20 tests per project, and an Enterprise plan whose terms are custom.

The published limits are volume and count caps: inference logs, members, projects, and tests. The reviewed sources publish no billing unit for the gateway or the guardrails, so a buyer cannot compare the gateway line's cost against a competing gateway from the public record.

The controls a regulated buyer needs sit behind the negotiated plan. Role-based access control, SAML single sign-on, on-premises deployment, compliance reports, data-retention control, and uptime commitments are all Enterprise-only, so the free plan is an engineering trial rather than a security evaluation.

The customer's own team sets the policy and owns the outcomes. Guardrails are created in the policy pages of the gateway's admin portal, everything is configured once with no redeploys and no code changes in the applications sending traffic, and the portal opens on a dashboard showing what the whole deployment is spending.

Hosting is regional, with data stored in the European Union or the United States, and Openlayer offers an air-gapped installation inside a customer's own network with no external connections for buyers who cannot send prompts outward.

Operational honesty is visible in the documentation. Output guardrails do not inspect streamed responses, and the vendor says so and explains why, telling teams to send requests without streaming when they need response-side screening. Tracing is fire-and-forget so it cannot slow or break a response.

Openlayer publishes a trust center at trust.openlayer.com and a security page listing SOC 2 Type II, GDPR, and HIPAA, monitored continuously with Vanta, with penetration tests run by independent security firms. A random control subdomain of the same domain did not resolve, so the trust host is genuine rather than a wildcard answer.

The enterprise access controls are the expected set: role-based access control, single sign-on through SAML and OIDC with identity providers such as Okta, directory sync, encryption in transit and at rest, and automated backups.

None of this blocks a replacement. A funded competitor selling into the same enterprises obtains the same attestations through ordinary preparation, so the posture is a credibility floor for the buyer rather than an obstacle for a rival.

The gateway's ecosystem position is defined by what it sits between. It speaks the provider APIs applications already use, can translate between OpenAI-style and Anthropic-style formats so one request runs against either, routes requests across approved providers on availability, performance, cost, or policy, and manages provider failover.

Upstream, Openlayer is a component in someone else's control plane as well as its own. Databricks announced the Unity AI Gateway partner ecosystem on 2026-06-17 and named Openlayer among the AI security integrations alongside CrowdStrike, Palo Alto Networks, Zscaler, and others, with Openlayer supplying observability, evaluation, and runtime guardrails into governed workflows.

That dual position cuts both ways. It puts Openlayer in front of buyers who never shopped for a gateway, and it makes the platform that carries it the party deciding whether the integration stays necessary.

Gabriel Bayomi, Rishab Ramanathan, and Vikas Nair founded the company in 2021 and hold the chief executive, technology, and product roles. SiliconANGLE reports independently that all three worked together on Apple's machine-learning efforts, and the company says that work covered models for Siri and Vision Pro.

The founding problem was evaluation rather than security. The company says teams at Apple kept rebuilding the same infrastructure to measure how models performed, that it started as Unbox and renamed to avoid conflicts, and that it grew from pre-deployment testing into production monitoring before adding governance and automated compliance.

Y Combinator lists a team of 24. The record carries no security engineering track record behind the guardrails, so a buyer weighing the controls has the documentation rather than the team's history to go on.