Lanbo Probe is the testing and guardrail layer for enterprise LLM apps. We attack your copilots, agents and chatbots with 4,000+ real-world exploits, score every release, and block what slips through in production. Works with any model. We don't train one.
Shipping safer AI at
Probe Red Team throws 4,000+ adversarial recipes at your app: multi-turn jailbreaks, indirect injection through documents and tool results, encoding tricks, and persona attacks. New techniques land in the library within 72 hours of going public.
Probe Evals turns quality and safety into a test suite that runs in CI. Groundedness, refusal accuracy, policy compliance and task success, with a pass or fail you can block a merge on.
Probe Guard sits between your app and any model as a drop-in proxy. It blocks injections, redacts PII, and rewrites ungrounded answers in a median of 18ms, then streams every verdict into Probe Trace for audits.
This is a real eval run animated at actual speed. Yellow passed, red failed. Red cells block the merge and land in your PR with the exact prompt, output and grader reasoning.
Point your SDK at Probe Guard. Keep your model, your prompts and your stack. Policies live in a YAML file next to your code.
from openai import OpenAI client = OpenAI( base_url="https://guard.lanbolab.pro/v1", # ← the only change default_headers={"x-probe-policy": "support-bot@v12"}, ) resp = client.chat.completions.create( model="gpt-4.1", messages=[{"role": "user", "content": user_input}], ) # resp.probe → {"verdict": "pass", "latency_ms": 17, "risk": 0.03}
import Anthropic from "@anthropic-ai/sdk"; const client = new Anthropic({ baseURL: "https://guard.lanbolab.pro/anthropic", defaultHeaders: { "x-probe-policy": "agent-tools@v3" }, }); const msg = await client.messages.create({ model, messages }); // blocked requests return a safe reply + msg.probe.verdict
policy: support-bot version: 12 input: - block: [prompt_injection, jailbreak, encoded_payload] - redact: [email, phone, iban, ssn] output: - require_grounding: { sources: kb/*, min_score: 0.8 } - rewrite: [policy_promise, competitor_mention] on_violation: { notify: "#ai-incidents", trace: true }
“Our security review used to take a quarter. With Probe’s report attached, it took nine days.”
“Probe found a tool-call injection in our agent on day one that three human pentests missed.”
“We swapped models twice this year. The eval suite told us exactly what broke before customers did.”
“18 milliseconds. Our users have no idea Guard is there, which is the whole point.”
“PHI redaction we could actually show our compliance team. That unblocked the launch.”
Every enterprise is putting LLMs in front of customers. Almost none can prove those systems behave. We are a small, opinionated team of security engineers and ML researchers building the missing layer: neutral, model-agnostic, and boring in all the right ways.
A green dashboard is not a safe model. Here are the seven failure modes we see most often when we audit customer eval suites, and how to fix each one.
When your agent reads email, tickets and web pages, every one of those sources becomes part of your attack surface. We need to start treating it that way.
A guardrail that adds a second of latency gets turned off. Here's the architecture behind Probe Guard's p95 under 30ms, and the trade-offs we made to get there.