Pervaziv AI adds local privacy and prompt safety controls to Cortex

Pervaziv AI said July 8 it is expanding Cortex with on-device local models that detect sensitive data, flag prompt-injection risks and securely deliver approved model versions inside developer workflows. The move is meant to cut latency, reduce remote inference costs and keep more AI safety decisions on the client side for enterprise software teams.

Why it matters: - Pervaziv AI is pushing AI safety and privacy checks closer to developers, which can reduce exposure of sensitive code, logs and credentials. - The company is also trying to lower latency and remote inference costs by handling routine safety decisions on-device instead of sending every check to a larger model. - The update strengthens Cortex as an enterprise control layer for secure software development, where privacy, governance and model behavior all matter.

What happened: - Pervaziv AI announced a major expansion of its on-device local model strategy for Cortex on July 8, 2026. - The release adds three capabilities: Cortex Privacy, Cortex Prompt Guard and Cortex Secure Distribution. - Cortex Privacy ships as cortex-privacy-1.1 for local sensitive-data detection and privacy-aware preflight scanning. - Cortex Prompt Guard ships as cortex-prompt-guard-1.2 for local prompt-injection and instruction-risk classification. - Cortex Secure Distribution handles private model delivery, versioning, integrity verification, provenance metadata, packaged product behavior and enterprise lifecycle management. - The company said the new work extends the direction set by Cortex 5.0 and Cortex-LLM-1.0, its first internally trained AI model for secure software development.

The details: - Cortex 5.0 introduced specialized AI behavior for security analysis, secure remediation, structured findings and secure agentic engineering workflows. - The local model strategy is designed to run inside VS Code and major browsers, including Chrome, Safari, Edge and Firefox. - The goal is to keep privacy and AI safety controls inside the tools developers already use. - Pervaziv AI said developer context can include credentials, tokens, private endpoints, database connection strings, cloud account identifiers, stack traces, logs, customer references, internal hostnames and operational data. - The same context can also include untrusted material from web pages, issue trackers, package metadata, pull request comments, documentation, generated text and copied logs. - Cortex Privacy is intended to identify sensitive content before it leaves the local environment. - Possible actions include warning the user, redacting sensitive spans, blocking unsafe sharing, routing the request differently or applying product-specific privacy behavior based on enterprise policy. - Cortex Prompt Guard is meant to detect prompt-injection and instruction-manipulation attempts before they influence AI behavior. - Prompt-injection risk can appear in ordinary developer surfaces such as web pages, package READMEs, dependency descriptions, pull request comments, issue tracker entries, copied logs and documentation pages. - Cortex Secure Distribution uses stable model versions, private delivery, integrity verification metadata, provenance metadata, runtime compatibility validation and release-level governance. - The distribution approach is meant to avoid direct runtime dependence on external model sources during normal product use. - The company said enterprises can get fewer setup requirements, fewer runtime access failures, stronger control over model availability and more predictable AI behavior. - Stable versions are meant to make privacy detection, prompt-risk classification and model behavior more repeatable across clients. - Versioning also supports governance, traceability, testing, promotion, rollback and reproduction of results.

Between the lines: - The release reflects a layered approach to AI, where local models handle fast privacy and safety checks while larger models handle deeper reasoning and remediation. - That architecture is meant to reduce overblocking and avoid unnecessary friction in developer workflows. - It also suggests Pervaziv AI sees enterprise adoption shifting from AI experimentation to AI operations, where control and reproducibility matter as much as capability. - The emphasis on secure distribution signals that model delivery itself is now part of the product, not just the model output.

What's next: - Pervaziv AI said the new local model capabilities are part of its continuing work across secure AI, DevSecOps, browser-based development, IDE-based development and enterprise AI governance. - The company said the roadmap continues to focus on model independence, on-device local inference, prompt-injection defense, privacy protection and secure agentic engineering. - Cortex will keep combining local controls with specialized models and larger reasoning systems depending on the task and the sensitivity of the context.

The bottom line: - Pervaziv AI is turning Cortex into a more complete enterprise AI control stack, with local privacy and safety checks built into the developer workflow instead of bolted on after the fact.