Security for AI agents
and the code that ships them.

OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Cohere, Mistral, and any custom provider via our adapter. Native hooks for LangChain, LangGraph, CrewAI, LlamaIndex, Semantic Kernel, Vercel AI SDK, Pydantic AI, and the OpenAI Agents SDK. Model- and framework-agnostic at the protocol layer.

p50 under 40ms, p95 under 120ms for inline audits. You can run Guardra synchronously (blocking) or asynchronously (fire-and-forget with later reconciliation). Fail-open mode keeps your agent running if Guardra is ever degraded — we never want to be the reason your product goes down.

Not unless you choose SaaS. Guardra runs in one of four modes: Guardra Cloud (multi-region, ephemeral enclaves), inside your VPC (data-plane stays in your cloud), on-prem, or airgapped. Source and traces are analyzed in memory and discarded — findings, not raw data, are what persist. Premium supports customer-managed keys (BYOK) and HSM.

They scan code. Guardra audits your agent AND your code, prioritizes by real exploitability, and ships the fix. In head-to-head trials we reduce false positives by 80%+ and cut median remediation time from weeks to under 15 minutes.

Every fix runs through a secondary adversarial-review model, is validated against a regression test we also author, and ships with a confidence score. You can require human approval on any severity, path, or repo via policy-as-code. 91% of our customers' fixes merge without human touch; the 9% that don't are the ones that should get reviewed.

Covered. Guardra scans every document entering your RAG index against our adversarial corpus (2M+ labeled examples). We also analyze retrieval outputs at inference time to detect injected instructions smuggled in via docs, emails, PDFs, or scraped web content.

Yes. 14-day pilot on every plan — full features, no credit card. Most teams find their first critical issue within 10 minutes of installing the SDK. For open-source maintainers: Guardra is free forever.

By audits/month, which covers every prompt, tool call, memory write, output eval, or code scan. Plans include generous monthly allowances (100K / 2M / unlimited). Overages are never billed automatically — you'll get alerts at 50%, 75%, 90% and can set hard caps.

Evidence exported on demand. Guardra maps its own controls and your audit findings into SOC 2 CC / ISO 27001 Annex A / PCI-DSS / HIPAA Security Rule / GDPR / EU AI Act Articles 9–15. Your auditor gets a signed, timestamped package — we've had customers close renewals in 48 hours.

Yes, on Premium. Kubernetes-native single-tenant deployment with offline update channel, customer-managed encryption keys, FIPS 140-3 validated crypto, and STIG-hardened reference manifests. FedRAMP Moderate in process; IL4/IL5 enclaves available for federal.

SOC 2 Type II, ISO 27001 / 27017 / 27018, PCI-DSS L1, HIPAA, GDPR, CSA STAR. Penetration-tested quarterly by external firms. Bug bounty via HackerOne. Full security reports and SBOM available at guardra.ai/trust.