Robust Intelligence

• Cisco sells the capability inside AI Defense rather than under the Robust Intelligence name, so investment in model validation follows Cisco’s roadmap for that product.
• Runtime guardrails are enforced through Cisco security and networking products, so a buyer that does not route AI traffic through Cisco infrastructure gets less of that delivery advantage.
• Cisco offers the enterprise edition’s algorithmic red teaming at no upfront cost in Explorer Edition, which could pull demand for paid model testing toward that entry point.
• Cisco disclosed no acquisition price and no revenue for the line, so its commercial scale before the deal rests on press estimates such as CTech’s $10 million figure.
• Model providers could ship validation and guardrails with their own models, which would narrow the opening for a separately bought testing product within three to five years.

Robust Intelligence serves enterprises that put AI models and applications into production and need to know how those models behave under attack. Cisco describes an automated, algorithmic assessment of a model’s safety and security vulnerabilities, kept current through AI Threat Research teams, and Cisco’s acquisition announcement describes protection for AI models across their lifecycle, from development to production.

The risks Cisco names are specific. Validation tests models with algorithmically generated prompts across 200 categories to find susceptibility to prompt injection and data poisoning, and supply chain scanning targets malicious model files that can allow arbitrary code execution. Cisco maps detections and tests to OWASP and MITRE ATLAS, and the data sheet adds the NIST AI Risk Management Framework, so a security team can check coverage against reference points it already uses. [s3, s4, s20]

The capability ships today as parts of Cisco AI Defense. AI Model and Application Validation tests models with algorithmically generated prompts across 200 categories and generates guardrails tailored to the vulnerabilities it finds in each model, and a customer can modify those rules to fit its industry or use case. An assessment can be initiated with a simple API call when a new model is entered into a registry.

Runtime protection screens prompts and responses, and Cisco’s data sheet says the guardrails cover API calls and communications between services and AI applications. Supply chain scanning checks model files, repositories, and the Model Context Protocol servers that AI agents use to reach tools, before those assets enter development or production.

A Cisco blog describes the engine behind the Robust Intelligence platform as three proprietary elements, algorithmic red teaming technology, an AI threat intelligence pipeline, and policy mappings. The team also published Tree of Attacks with Pruning, an automated method for jailbreaking models that needs only access to model inputs and outputs without access to model internals. [s3, s20, s2, s10, s14]

Inside Cisco the line reaches buyers through Cisco’s own products. The acquisition announcement describes delivering AI security processing into existing data flows by inserting it into Cisco security and networking products, and Cisco says it enforces AI security at the network level without the need for agents or libraries.

Cisco also offers the same testing engine through Explorer Edition, a self-service option carrying the enterprise edition’s algorithmic red teaming at no upfront cost. Cisco presents it as a way for AI builders to red team their applications before scaling to an enterprise solution. That reach and that entry point belong to Cisco, and the cited record documents no head-to-head technical comparison with another vendor. [s4, s2, s12, s20]

Before the acquisition the company sold a standalone platform. PayPal and Robert Bosch were customers of the Robust Intelligence Model Engine, according to SiliconANGLE. Customers could deploy the platform in the cloud or on premises. CTech reported approximately $10 million in revenue and 70 employees at the sale.

The acquisition is the clearest traction event in the record. Cisco announced its intent on August 26, 2024, posted a completion update dated September 24, 2024, and its product page dates the acquisition to October 2024.

Distribution now runs through Cisco. Cisco says the validation engine behind Explorer Edition is trusted by Global 2000 customers. [s9, s7, s4, s19, s21]

The founding pair is well documented. TechCrunch reports that Yaron Singer, a tenured professor of computer science and applied mathematics at Harvard, founded the company with his former student Kojin Oshiba, and CTech describes a professor who became a Harvard faculty member at 33 and earned tenure by 40 after publishing machine learning and artificial intelligence research. BankInfoSecurity reports that the company was founded in 2019.

The research record continues into the product era. Tree of Attacks, accepted at NeurIPS 2024, lists Robust Intelligence authors alongside Yale collaborators. After the deal, Cisco described Singer as vice president of AI and engineering. Cisco says its Foundation AI researchers joined through the acquisition and released an open-source reasoning model for security applications. [s8, s7, s10, s18, s14]

The line leans on shared standards for assurance. Cisco maps all detections and tests to OWASP and MITRE ATLAS, the data sheet adds MITRE ATLAS, the OWASP Top 10 for LLMs and the NIST AI Risk Management Framework, and validation generates vulnerability reports mapped to industry and regulatory standards.

Deployment choices speak to buyers with data-handling constraints. In virtual private cloud and on-premises AI POD deployments, Cisco says application, model and agent data stay within the customer’s environment. Cisco receives only metadata needed to manage those deployments. That option lowers a procurement objection about where inference traffic travels, and it is an adoption aid rather than a lock. [s4, s20, s3]

Add analyzed competitors to compare them side by side with Robust Intelligence.

The line serves enterprises that build or run AI applications. Cisco points validation at teams that download models and data from public repositories, including Hugging Face and GitHub, and points runtime guardrails at teams running AI applications and agents in production.

The named buyers are large enterprises. SiliconANGLE reported PayPal and Robert Bosch as customers of the Robust Intelligence Model Engine. Co-founder Yaron Singer told CTech about a client at a major U.S. bank. Since the deal Cisco says the validation engine is trusted by Global 2000 customers, which is a statement about the engine’s reach rather than named accounts for the line.

The capability is specific and hard to build. Validation tests models with algorithmically generated prompts across 200 categories, generates guardrails tailored to each model’s vulnerabilities, and scans model files, repositories and Model Context Protocol servers for hidden risks before they enter development.

The AI advantage Cisco describes is continuous updating. The assessment is kept current through AI Threat Research teams, and the data sheet says AI Defense is continuously updated with attack data from Cisco’s AI threat research team and Talos. A Cisco blog calls the algorithmic red teaming technology, the AI threat intelligence pipeline and the policy mappings proprietary.

The team’s published research includes the red-teaming work. Tree of Attacks with Pruning, accepted at NeurIPS 2024, is an automated method for generating jailbreaks that needs only black-box access to the target model, and its author list carries Robust Intelligence. Cisco separately calls the algorithmic red teaming behind the platform proprietary.

Go-to-market runs through Cisco. The acquisition announcement describes inserting the technology into Cisco security and networking products, and Cisco says the validation engine behind Explorer Edition is trusted by Global 2000 customers.

Two entry points now exist. Enterprise buyers request a demo of AI Defense, and AI builders can start with Explorer Edition, which offers the enterprise edition’s algorithmic red teaming at no upfront cost. The cited sources do not report how many Explorer Edition customers upgrade to paid deployments.

Before the deal the line sold a standalone platform, and SiliconANGLE reported pre-acquisition use of the Robust Intelligence Model Engine by enterprises such as PayPal and Robert Bosch.

Cisco’s AI Defense product page carries no price and invites buyers to request a demo, which is the path the page gives an enterprise buyer.

The visible offer is Explorer Edition. It carries the same algorithmic red teaming capabilities as the enterprise edition at no upfront cost, so a developer can test models and agents before any purchase. The cited pages document no charging unit for the enterprise edition.

Delivery sits inside Cisco’s network enforcement. Cisco enforces AI security at the network level without the need for agents or libraries, which decouples AI development from security, and the guardrails cover API calls and communications between services and AI applications.

Operations favor teams that already automate. Validation can be built into existing CI/CD workflows, an assessment can be initiated with a simple API call when a model enters a registry, and findings reach security consoles such as Splunk through integrations. In virtual private cloud and on-premises AI POD deployments, Cisco says application, model and agent data stay within the customer’s environment. Cisco receives only metadata needed to manage those deployments.

Standards mapping carries most of the trust argument. All detections and tests are mapped to industry and regulatory standards like OWASP and MITRE ATLAS, the data sheet adds the OWASP Top 10 for LLMs and the NIST AI Risk Management Framework, and validation generates reports mapped to industry and regulatory standards.

That alignment is shared vocabulary rather than a gate. The data-residency option, where customer data stays in the customer’s environment in VPC and AI POD deployments, answers a separate procurement question about where AI traffic travels.

The line is a component of a larger platform. Cisco scans public model files through partnerships such as Hugging Face using Talos threat intelligence and ClamAV, connects findings to security consoles such as Splunk, and has added a developer-ready integration with NVIDIA NeMo Guardrails, an open-source framework.

Its neighbors inside Cisco keep growing. The February 2026 update brought AI supply chain governance and runtime protections for agentic tool use, and the Foundation AI team, which joined through the acquisition, released an open-source reasoning model for security applications. The Robust Intelligence team also helped establish AI security standards through collaboration with MITRE, NIST and OWASP, which ties the line to the ecosystem it is measured against.

The founders brought verifiable domain pedigree. TechCrunch reports that Yaron Singer, a tenured professor of computer science and applied mathematics at Harvard, founded the company with his former student Kojin Oshiba, and CTech describes a career that reached a Harvard professorship at 33 and tenure by 40. BankInfoSecurity reports that the company was founded in 2019.

The team stayed after the sale. Cisco described Singer as vice president of AI and engineering when it announced the Gartner Cool Vendor mention. Cisco says the Foundation AI team joined through the Robust Intelligence acquisition. Cisco has kept shipping on the line since, expanding AI Defense at Cisco Live EMEA in February 2026 and opening the self-service edition in March 2026.