How it works

Inspected at every hop — in, and out.

A deterministic path every agent request takes — five stages, one service, entirely inside your network. Fail-open on infrastructure, sub-second added latency.

Step 01

Identify

Auth, rate & budget.

Step 02

Inspect

Injection, poisoning, tool abuse.

Step 03

Route

Provider & failover.

Step 04

Restore

De-tokenize, egress scan.

Step 05

Record

Crypto audit → store.

STAGE 01 · IDENTIFY

Who is asking, and can they?

Every request is authenticated, rate-limited, and budget-checked at the door — before a single token is spent. Identity flows through to per-agent policy and the audit record.

  • API-key, mTLS, or SSO-backed agent identity
  • Per-tenant and per-agent rate & budget enforcement
  • Deterministic policy evaluation in the hot path
agent: support-copilotauthed
rate: 142 / 600 rpmwithin
budget: $820 / $1,500ok
STAGE 02 · INSPECT

Every surface, every attack — in real time.

Each surface an agent touches carries its own attack: prompt injection and jailbreaks, poisoned documents through RAG, malicious or 'toxic-combination' tool and MCP calls, memory poisoning across a session, unverified trust between agents, and data exfiltration on the way out. SHASHU by Trustgate AI inspects every one of them through its detection guard, which learns each agent's normal behavior. Deterministic regex and DLP are the fast baseline underneath; the detection guard catches the behavioral attacks they can't see.

  • Prompt injection, jailbreaks & RAG poisoning — detected
  • Malicious & toxic-combination tool / MCP calls — scoped
  • Self-healing scrutiny: NORMAL → ELEVATED → STRICT, automatically
  • Benchmarked to OWASP LLM Top 10, MITRE ATLAS & NIST AI RMF
prompt · injection attemptblocked
rag · poisoned documentquarantined
tool · call scopedallowed
STAGE 03 · ROUTE

To the right model, every time.

Once a request clears inspection, SHASHU routes it to the chosen provider with automatic failover — OpenAI, Anthropic, Bedrock, Azure, or your self-hosted models — behind one OpenAI-compatible endpoint.

  • Multi-provider with health-aware failover
  • Drop-in OpenAI-compatible — no app changes
  • Self-hosted & GPU model targets supported
primary: anthropic / claudehealthy
failover: azure openaistandby
self-hosted: llama-guardlive
STAGE 04 · RESTORE

Clean on the way out, too.

The response is de-tokenized to restore real values, scanned for PII leakage and exfiltration, and run through egress forensics — so what leaves the model is as governed as what entered it.

  • Reversible PII tokenization restored in-response
  • Output scan for leakage & data exfiltration
  • Egress forensics on every response
detokenize · 4 values restoreddone
egress scan · no PII leakclean
data egressed · 0 bytesenforced
STAGE 05 · RECORD

A trail you can actually prove.

Every request is written to a cryptographically chained, tamper-evident audit trace — then streamed and stored where you choose. Evidence your auditors and procurement team can verify, not just trust.

  • Cryptographically chained audit records
  • Stream to SIEM · store in your own bucket
  • Export-ready evidence per framework
audit chainLive
Tamper-evident · verified chain
14:02:47req_8f21 · support-copilot#a3f9…c1
14:02:46req_8f20 · billing-agent#7b2e…9d
14:02:45req_8f1f · rag-indexer#e04c…2a
14:02:44req_8f1e · ops-assistant#11bd…7f
Performance & failure profile

Built to sit in the hot path.

One service in your VPC, designed so security never takes your agents offline.

<1s
added p95 latency
0 B
data egressed
0
surfaces inspected
Fail-open
on infrastructure
Fail-open on infrastructure, fail-closed on threatsNothing leaves your network

See the pipeline on your own traffic.

Deploy the open core in three commands, or book a walkthrough with our team.