Obvix-Proxy
Obvix-Proxy sits between users and language models to intercept every exchange, make context-aware decisions, and block or edit harmful content mid-flight without retraining the underlying model.
Average latency
~58 ms
Measured on consumer RTX 4060 GPU
Status
Under testing
Preparing for general release after refinements
Focus
Context-aware filtering
Rule checks + AI reasoning + tool lookups
Architecture
Layered guardrails, operator friendly
The proxy intercepts every request + response, adds context (risk levels, compliance flags), and streams it to your dashboards. You choose when to warn, block, or escalate.
Dialogue intercept
The proxy inspects every user prompt and model reply, editing or blocking harmful content before it reaches either side.
Context reasoning
Rule checks work alongside AI understanding to distinguish benign mentions from real policy violations.
Tool + data verification
When something is unclear, Obvix-Proxy calls external tools or knowledge bases to make a more informed decision.
Tool-call containment
Function / tool invocations run through allowlists so jailbroken prompts can't trigger harmful automations or systems.
Rollout path
Shadow mode
Drop-in proxy sits between client + LLM, mirroring traffic while surfacing incidents.
Auto-evals
Telemetry feeds nightly evals. Failures open issues in the shared notebook.
Progressive enforcement
Start with warn-only, graduate to block/edit once humans approve.
Governance sync
Policy teams get exportable evidence packs for regulators.
Live log trace
Agent attempted admin-level SSO impersonation without ticket reference; proxy blocked privilege escalation.
{
"userId": "usr_5021",
"reason": "debug-session"
}Preflight scan
- Verdict:
- Hard block
- Signal:
- Privileged action without JIRA key
- Action:
- Suppressed call; returned guidance to open change request.
Postflight scan
- Verdict:
- n/a
- Signal:
- Execution never dispatched
- Action:
- Security event logged with severity=high.
Telemetry console
Observation deck built-in
Every intervention, warning, and escalation is logged with rich context: conversation snapshots, persona tags, mitigation actions.
Every intervention is logged with the surrounding dialogue
Notes explain whether rules or AI reasoning triggered the block
Tool/database lookups are recorded for auditing
Summaries help teams brief leadership and regulators
Latency profile
Fast enough for chat, rigorous enough for ops
Internal tests show an average latency of ~58 ms on commodity RTX-4060 hardware, so conversations stay natural while guardrails run.
Because the proxy works independently from the base model, we can keep tuning rules and tool lookups without retraining your LLM.
Status: under testing and refinement-we’re expanding scenarios before general release.
Red teaming
Adversarial testing built into the proxy
Obvix-Proxy doesn't just block — it hunts. Red-team modules run persona vectors, exploit chains, and dataset audits through the same proxy pipeline, so you find weaknesses before attackers do.
Layered jailbreak sequences that escalate across multi-turn conversations, surfacing which filters collapse and what data spills when they do.
Attack personas that coax models into unsafe modes without obvious jailbreaks — scoring each conversation on impact x likelihood for crisp prioritisation.
Dataset integrity audits that flag corrupted embeddings, prompt-leaked secrets, and hallucination hotspots by comparing normal versus adversarial recall.
Plays well with
All major AI workflows
Agentic orchestrators, conversational copilots, evaluation harnesses, and classic RAG systems all sit behind the same proxy-no need to rewire when you expand coverage.
Tool-call circuit breaker
We even stop AI agents from calling harmful tools when bad actors attempt a jailbreak-the proxy inspects every function call before it reaches production systems.
Provider coverage
Mix and match models
Plug Obvix-Proxy into whichever foundation model makes sense today-OpenAI, Anthropic, Vertex AI, Azure, AWS Bedrock, OpenRouter, or your own OSS stack.