SandboxAQ has added two products to its Switch AI platform. Switch Trust finds a company’s AI agents, tests them for attacks such as prompt injection, and blocks unsafe behaviour. Switch ROI tracks each agent’s spending against outcomes. Both are free for up to five agents.
SandboxAQ has launched two tools to help companies check that their AI agents are safe and worth what they cost, it said on Monday. Switch Trust and Switch ROI are part of its Switch AI platform, which launched in August.
Switch Trust first lists the AI agents, models, and tools running in a company’s systems, including ones its security team may not know about. It then scans them for security flaws and tests them against attacks such as prompt injection, where hidden instructions trick an agent into misbehaving.
Once an agent is live, Switch Trust checks what goes in and out of it, the company said. It can block or hide leaked passwords, personal data, and other unsafe content, and it keeps a record of what each agent did.
Switch ROI tracks how much each agent costs and links that spending to results, such as support tickets closed or tasks automated. It also suggests savings, such as moving work to a cheaper AI model or cutting tools an agent never uses.
Both products work with agents from any vendor, SandboxAQ said. They are free for up to five managed agents, according to a company blog post, which did not list paid prices.
Mohammed Aboul-Magd, who runs Switch AI, said the company had seen AI spending meant to replace a new hire end up costing more than a full-time employee. Mount Sinai West’s MSW Lab is using Switch Trust while its doctors build medical software with AI agents, SandboxAQ said.
The launch comes as many companies struggle to get value from agents. Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. AWS has also said almost 90% of Amazon’s early agent prototypes never shipped.
Security is a growing worry too. Google, JPMorgan, and others have fixed the same flaw in MCP, a common standard for linking agents to tools.