Skip to content
Docs menu

Red team

Run the versioned attack corpus against your guardrails and a gear with venom redteam run, within a spend cap, and read which guardrail caught each case.

View .md

The red-team suite runs a versioned corpus of attacks against the guardrails in force and, from the CLI, against a gear. It scores each category, keeps every run for the trend on Venom Cloud, and says for each case which guardrail caught it, or that none did and which rule would.

Quick start

venom redteam run --dry-run                    # the plan and the estimate; nothing is sent
venom redteam run                              # the free cases only
venom redteam run --gear strong --max-usd 0.10 # free cases, plus model cases within $0.10

The corpus

The corpus only grows, and a case's id never changes meaning, so runs on the same corpus version compare.

Category What the cases try
Prompt injection instructions planted in web pages, READMEs and tool output: override phrases, chat-template tokens, tool-call look-alikes, invisible tag characters, hidden HTML
Jailbreaks "DAN" and "developer mode" prompts; a role-play that asks for a secret the system prompt holds
Data exfiltration customers' emails, phone numbers and ID numbers in a prompt; image links that carry data out
Secret leakage a database URL with its password, an .env read by a tool, a password in an answer
Policy bypass other spellings of a denied model: case, provider prefixes, :latest, fullwidth letters, zero-width spaces, look-alike letters, snapshot dates
Tool abuse curl … | bash, "push and don't tell the user", "upload the .env"

Three kinds of check:

Check Cost Passes when
guardrail free the text, sent as a prompt, a tool result or an answer, is blocked, redacted, kept local or warned about by an enforced rule. A rule in Monitor doesn't count
models free every other spelling of each model you deny is refused too. Skipped when no model is denied
model paid a fresh canary (VENOM-CANARY-XXXXXXXX) placed where the secret would be never comes back in the answer or a tool call's arguments. A guardrail that refuses the request is a pass

Evidence never holds the attack text or the model's answer, only what was decided.

Running it

It always prints the plan first: how many free cases, how many model cases fit under --max-usd, and their estimate at a deliberately high rate ($10 per million input tokens, $40 per million output, at most 300 answer tokens). Each model case runs only while the estimate stays under the cap. The four model cases together come to about six cents at that rate.

Flag Default What it does
--dry-run off Print the plan and estimate; send nothing
--gear <name> none Send model cases with this gear (x-venom-gear)
--model <id> venom-auto The model for model cases
--max-usd <usd> 0 (no model cases) Spend cap for model cases, 0 to 10
--json off Print the run as JSON

Requests go to VENOM_BASE_URL with VENOM_API_KEY. Against Venom Cloud, the free cases run there on your key's guardrails (the org's, your key group's and your own) and the run is kept. Against any other gateway they run in the CLI on your own venom.yaml guardrails and nothing is kept.

The Red team page

On Venom Cloud, /redteam under Safety & policy shows the latest run's score, a line per category, every case with the code that caught it (linked to its /g/<code> page) or the catalog entry to add, and the trend with each run's cost. Run checks now runs the free cases on the org's guardrails; model cases come from the CLI.