DUBLIN, Calif., October 15, 2026: Safeguard, the AI-native software supply chain security platform, today opened early access to Safeguard Guard, runtime guardrails that run inside an application. Guard is built to install as one line of middleware in Node.js, Python, Java, Go and .NET applications and enforces the guardrails a team defines on the Safeguard Policy page. Its SDK packages are rolling out to registry.safeguard.sh over the coming days.
A firewall in front of an application sees requests. Guard works at the library and framework level inside the application, where it can see what a request turns into: the SQL query it becomes, the outbound call it triggers, or the prompt sent to a model.
What Guard checks
- SQL injection, at query time. Guard checks whether request input changed the structure of a query as it runs. A parameterised query never triggers the check.
- Command injection, path traversal and unsafe code paths. Request input reaching a shell command or a code evaluator, file paths that leave an allowed root, and unsafe deserialization.
- SSRF and outbound calls. Outbound connections are checked against the IP addresses a hostname actually resolves to, and cloud metadata addresses can be blocked.
- LLM calls. Prompt injection, secrets and personal data in prompts and outputs, model allow lists, and token and cost limits.
- MCP tool calls. Allow, deny or monitor tools by name, and check tool arguments and results.
Built to be safe to switch on
Every guardrail starts in Monitor mode, which records what Guard would have done without changing any traffic. Moving to Enforce takes an explicit decision by a person. Guard fails open, so if it cannot decide, traffic goes through and the event is recorded. A kill switch turns every check off without a code change. Each decision shows where a check fired, such as the library call and the request field involved, without recording payload values.
Guard's in-process checks are designed for sub-millisecond overhead and make no network calls while handling a request.
One middleware line per framework
Once its SDK packages are available, Guard installs with a single middleware line:
- Express:
app.use(guard()) - FastAPI:
app.add_middleware(SafeguardGuard), with middleware for Django and Flask - Spring Boot: a starter that registers Guard automatically
- ASP.NET Core:
UseSafeguardGuard() - Go:
guard.Middleware(handler)
Zero-code options will also be available: a Node.js --require flag, a Python run wrapper, a Java -javaagent and a .NET startup hook. Guard can also run as a proxy or sidecar for services that cannot be changed.
Guard is distributed only through Safeguard's private registry, registry.safeguard.sh, with a pull token, and its packages are rolling out there over the coming days. It is not published to public package registries, and the registry never passes Guard's package names through to a public source, so a look-alike public package cannot be served in its place. Every release will be signed and carry an SBOM and SLSA provenance.
"A firewall at the edge sees a request. It can't see the query that request turns into, or what your application sends to a model," said Hritik Kumar Sharma, founder and CEO of Safeguard. "Guard runs where the input is actually used, so it can judge it by what it would do. It starts in Monitor mode and fails open, because a security tool shouldn't be the reason a service goes down. The goal behind everything we build is to shorten the time between finding a problem and stopping it, from days to hours, then minutes, and eventually seconds and milliseconds. Guard is how a team can put a guardrail in place while the fix is still on its way. It's early, and we want the first teams running it to tell us where it falls short."
Safeguard has published an educational series alongside this announcement, including a guide to what library-level protection catches that a WAF can't, at safeguard.sh/resources/blog/runtime-guardrails-inside-your-app-what-a-waf-cant-see.
Availability
Safeguard Guard is in early access. Its SDK packages are rolling out to registry.safeguard.sh over the coming days. Guard has not yet been tested by customers, and Safeguard is inviting its first early-access teams. Pricing is available on request. To request access, book a call at safeguard.sh/book-a-call.
About Safeguard
Safeguard is the AI-native software supply chain security platform. It finds risk across code, dependencies, containers and AI, and fixes it autonomously with tested pull requests. Learn more at safeguard.sh.
Media contact: press@safeguard.sh
