AI Security Vendors in 2026: Who Owns Protect AI, Lakera, Robust Intelligence and the Rest, and What They Do

Between 2024 and 2026, most of the first wave of AI security start-ups were bought by established security companies. If you interview at one of those acquirers, or are asked to evaluate "an AI firewall", you need to know who now owns what. This page is a dated map. It is not a ranking, and vendors' claims about their own detection rates are left out.

Who owns what (November 2026)

How these products work

Deployment shapes differ, and the shape decides what a product can see:

• API or SDK calls from your application code, like Lakera Guard's /v2/guard endpoint. These see exactly what the application sends, and miss anything it forgets to send.
• Network or gateway interception: Cisco AI Defense describes runtime guardrails embedded in the network "without the need for agents or libraries", and Prisma AIRS offers network as well as API intercept. These see traffic without code changes, but need TLS visibility and understand application context less well.
• Browser and endpoint controls for employees' use of public AI tools, as in Prompt Security and Cato AI Security.
• Scanners and red-team services that run before deployment rather than inline.

How to use it: an evaluation plan

Day one: write down what you need to see and stop

List your AI assets (applications, agents, MCP servers, employees' tools) and the decisions you want enforced. A platform that only screens chat prompts does nothing for an agent's tool calls, and a network intercept does nothing for traffic it cannot decrypt.

Week one: test with your own scenarios

Run every candidate against the same test set, built from your own use cases:

• known attacks
• legitimate requests that look suspicious
• indirect injection planted in documents and tool outputs
• multilingual and encoded inputs

Measure misses, false positives and added latency. Read the documentation for what each product says it does not cover. Ask where your data is processed and stored, especially for products that see full prompts and responses.

Every day: integrate and keep evidence

Send detections to your SIEM, define who responds to an AI incident, and re-run the evaluation when models, prompts or the product change. After an acquisition, watch the roadmap: products are often renamed, merged into suites or repriced.

Where they stop

• Classifiers are not guarantees. Runtime protection reduces risk; architecture (least-privilege tools, approvals, no secrets in context) limits the damage when it misses.
• Visibility depends on placement. An API-based guard misses calls the app does not route through it; a network guard misses encrypted or local traffic.
• Vendor numbers are claims. Detection rates in marketing come from the vendor's own test sets.
• The market moves quickly. Check ownership and product names on the date you read this.

How it comes up in interviews

• "Our CISO wants to buy an AI firewall. How would you evaluate it?" Use your own test set, documented gaps, placement, data handling, latency and incident integration.
• "What changed in the AI security market in 2025?" Consolidation into platform vendors, and what that means for buyers.
• "Build or buy guardrails?" Open-source and cloud guardrails against vendor platforms, by coverage, effort and evidence.

Sources

Checked on 13 November 2026: Palo Alto Networks press release on completing the Protect AI acquisition (22 July 2025) and Prisma AIRS documentation; Check Point press release on acquiring Lakera and Lakera documentation; Cisco blog on acquiring Robust Intelligence and the Cisco AI Defense announcement (January 2025); SentinelOne SEC filings on the Prompt Security acquisition; F5 blog on AI Guardrails and F5's FY2025 annual report; Help Net Security and Cato Networks announcements on Aim Security and Cato AI Security; Vanta announcement on acquiring Pillar Security (9 November 2026); HiddenLayer Series B announcement (2 September 2026); SecurityWeek reports on Noma Security and Mindgard funding; SiliconANGLE report on Lasso Security (2 September 2026).