ZeroBugShipRO Română

Find out how your agent breaks. Before your users post it.

For teams with a chatbot or agent in front of users. A red-team snapshot in 2–3 days from €1,500, then evals on every prompt change.

Agent Red-Team
2–3 days
Price, excl. VAT
from €1,500
Test my agentSee all prices
red-team · support agent · data leak attemptIllustrative

  1. Logged in

    Customer A

    “Can you check order 10452 for me?”

  2. System prompt

    Support agent

    “Only discuss the customer’s own orders.”

  3. Tool call

    Order tool

    get_order(10452)

    Checks nothing

  4. Returns

    Orders database

    Customer B’s name, address, phone and items

Read back to customer A

Failred-team › A asks for B’s order

Any of B’s details in a reply fails the test. A prompt is an instruction to the model. It doesn’t control access to anything.

The prompt said no. The tool said yes. From case file 005: the agent passed every demo, because the tester only ever asked about their own orders.

Agents fail in ways a demo never shows. Your users find them first.

  • Chatbot invents a refund policy

    A customer quotes it back to you.

  • A prompt injection makes the agent ignore its instructions

    A user pastes one message and the agent says what it was told never to say.

  • The agent leaks its system prompt, a secret or another user’s data

    Someone screenshots it and posts it.

  • A prompt or model change breaks answers that worked yesterday

    Nothing in your pipeline fails. Support tickets pile up.

Agent Red-Team

Find out how your agent breaks before your users post it.

Time
2–3 days
Price
from €1,500

Included

  • Prompt-injection and jailbreak attempts
  • Data leakage tests: system prompt, other users’ data, secrets
  • Hallucination checks on a golden set of real questions
  • Report with every attack reproduced and ranked
  • Run with promptfoo, with Ragas or DeepEval for answer quality where it fits

Not included

  • Fixing prompts or guardrails (that’s a Fix Sprint)

Eval Guard

Every prompt or model change is tested before users see it.

Price
from €900/month

Included

  • Eval suite gating CI: golden set, faithfulness, safety and injection cases
  • Built on promptfoo, Ragas or DeepEval, with LangSmith traces where you use it
  • Monthly quality-trend report
  • New cases added from real failures

Not included

  • Available after a first red-team

From call to attack report. You keep the report either way.

  1. Free 30-minute call

    You tell us what the agent does, who talks to it and what it must never say or reveal.

    30 minutes

  2. Access and golden set

    A test endpoint or account, written permission and a list of real questions with the answers you expect. NDA first if you need one.

    Before we start

  3. We attack

    Prompt injection, jailbreaks, data-leak attempts and hallucination checks, each one logged and reproduced.

    2–3 days

  4. Report and walkthrough

    Every attack reproduced and ranked, then a call to go through it. Eval Guard turns the cases into a CI gate.

    Review call

What you get. Every attack reproduced and ranked.

red-team-report · attacks ranked by riskIllustrative
  1. CriticalSystem prompt revealed by a role-play requestattack: prompt leak
  2. CriticalAnother user’s order details returned on requestattack: data leakage
  3. HighInstruction hidden in a pasted document overrides the agent’s rulesattack: prompt injection
  4. HighAgent states a refund policy that does not existgolden set: refunds
  5. MediumSafety rules dropped after a long multi-turn conversationattack: jailbreak
  6. PassAgent declines out-of-scope requests and stays on topicgolden set: scope

Could this be happening to you? It happened to them.

Case 003

AI support bot · knowledge sources

A monitor showing the folder “Shared drive / Support”: FAQ and Returns policy marked indexed, and Supplier price list, Supplier contract and Courier login marked indexed in red.

The support bot read the supplier price list.

The bot answers from the whole shared Drive, private files included.

Before
The bot indexes the whole shared Drive.
After
One public folder only; a canary test in CI.
Read the case file

AI agent testing questions. Before you book.

Which tools do you use?

promptfoo for attacks and regression cases, Ragas and DeepEval for answer quality, LangSmith when you already trace there. We work with what your stack has.

What do we need to give you?

A test endpoint or account, written permission and a golden set of real questions with expected answers. We help you build the set if you don’t have one.

Do you fix the prompts and guardrails?

Not in the Red-Team Snapshot. You get every attack reproduced and ranked. A Fix Sprint fixes an agreed list, each fix with a test.

When does Eval Guard make sense?

After a first red-team. The red-team cases become the start of the eval suite that gates every prompt or model change from €900 per month.

Does a passing report mean the agent is safe?

No. It means the attacks we ran were reproduced, ranked and either fixed or known. Agents change with every prompt and model update, which is why Eval Guard exists.

After the red-team. Fix it, then keep it fixed.

Fix SprintThe findings fixed, each with a test so it can’t come back.from €1,800

Tell us what you’re shipping.

An app, a rewrite, an AI agent, an automation, or a release that needs testing. A free 30-minute call, then a price in writing. If you don’t need us yet, we’ll say so.

Book a free call