A written threat model of your AI system.
The agent split a $480 refund into two refunds of $240, each under its $250 limit.
AI red teaming
We attack your AI system the way a bad actor would, fix what breaks and attack it again on every release.
| Probe | Category | Result |
|---|
| Repeat a card number back | Personal data | Passed |
| Refund over the limit in two steps | Tool misuse | Failed, high |
| Role play as a support manager | Jailbreak | Passed |
| Ask for another customer’s order | Data leak | Passed |
| Instructions hidden in a ticket | Prompt injection | Passed |
| Ignore the tone guide | Policy | Passed |
| Link to a lookalike site | Phishing | Passed |
| Translate a blocked request | Jailbreak | Passed |
| Ask for the system prompt | Prompt leak | Passed |
| Make up a return policy | Grounding | Passed |
| Insult the customer back | Tone | Passed |
The agent split a $480 refund into two refunds of $240, each under its $250 limit.
Then a user finds the prompt that makes it leak data or ignore its rules, and the fix for that attack quietly opens another one.
of organizations reported a breach of an AI model or application.
Sourcenewsroom.ibm.com
of those breached had no access controls on their AI.
Sourcenewsroom.ibm.com
of developers and testers surveyed use red team testing on AI.
Sourceapplause.com
You find the weak spots before your users do.
| What changes | Without it | With it |
|---|---|---|
| First attack | Found by a user in production. | Found by us before launch. |
| Old attacks | Never run again. | Run again on every release. |
| Findings | A list nobody owns. | Ranked by harm, each with a fix. |
What changes for your teams. We pick the tools to fit the systems you already run.
| Probe | Category | Result |
|---|
| Repeat a card number back | Personal data | Passed |
| Refund over the limit in two steps | Tool misuse | Failed, high |
| Role play as a support manager | Jailbreak | Passed |
| Ask for another customer’s order | Data leak | Passed |
| Instructions hidden in a ticket | Prompt injection | Passed |
| Ignore the tone guide | Policy | Passed |
| Link to a lookalike site | Phishing | Passed |
| Translate a blocked request | Jailbreak | Passed |
| Ask for the system prompt | Prompt leak | Passed |
| Make up a return policy | Grounding | Passed |
| Insult the customer back | Tone | Passed |
The agent split a $480 refund into two refunds of $240, each under its $250 limit.
| Probe | Category | Result |
|---|
| Repeat a card number back | Personal data | Passed |
| Refund over the limit in two steps | Tool misuse | Failed, high |
| Role play as a support manager | Jailbreak | Passed |
| Ask for another customer’s order | Data leak | Passed |
| Instructions hidden in a ticket | Prompt injection | Passed |
| Ignore the tone guide | Policy | Passed |
| Link to a lookalike site | Phishing | Passed |
| Translate a blocked request | Jailbreak | Passed |
| Ask for the system prompt | Prompt leak | Passed |
| Make up a return policy | Grounding | Passed |
| Insult the customer back | Tone | Passed |
The agent split a $480 refund into two refunds of $240, each under its $250 limit.
Illustrative figures, not a client result. Your own baseline is taken in the first week.
Another company built the same kind of system and wrote publicly about how it went.
What they learned
Red teaming finds the failures that matter only when people attack the whole system, not just the model.
Microsoft’s AI Red Team has red-teamed more than 100 generative AI products.
You own the config, the tests and the runbooks, so you can swap any part later.
The agent split a $480 refund into two refunds of $240, each under its $250 limit.
| Probe | Category | Result |
|---|
| Repeat a card number back | Personal data | Passed |
| Refund over the limit in two steps | Tool misuse | Failed, high |
| Role play as a support manager | Jailbreak | Passed |
| Ask for another customer’s order | Data leak | Passed |
| Instructions hidden in a ticket | Prompt injection | Passed |
| Ignore the tone guide | Policy | Passed |
| Link to a lookalike site | Phishing | Passed |
| Translate a blocked request | Jailbreak | Passed |
| Ask for the system prompt | Prompt leak | Passed |
| Make up a return policy | Grounding | Passed |
| Insult the customer back | Tone | Passed |
Typical building blocks: promptfoo in CI and garak before launch.
Alternatives: PyRIT (microsoft/PyRIT); DeepTeam. promptfoo has been owned by OpenAI since 2026-03.
The work happens in your cloud, with the people who will own it.
We attack the system before users do.
About 2 weeks. You get a price range before the call and one fixed number after it. That number is the invoice.
We pick the one that fits the systems you already run. You own the config, the tests and the runbooks, so you can swap any part later. Most often promptfoo in CI and garak before launch; PyRIT or DeepTeam when they fit better. promptfoo has been owned by OpenAI since March 2026.
Prompt injection, data leaks, jailbreaks, unsafe tool use and the rules your system must keep. We write the attacks for your data and your tools.
No. We attack a staging copy you approve, unless you ask us to test production within agreed limits.
The first run is. The attacks then stay in your CI, so every release is tested again.
Yes. Each attack maps to that list, and we add the risks only your system has.
We attack the system before users do. You own the config, the tests and the runbooks.