The 2026 red-team tooling stack
runtime defense: Llama Guard + Prompt-Guard
user input
Prompt-Guard-86M
Llama Guard 3/4
target LLM
model under deployment
14 MLCommons hazards watched
output
Llama Guard 3/4
category scores
user
response delivered
or refusal issued
regression: Garak nightly
probes
hallucination / leak / injection / jailbreak
static / dynamic / adaptive
detectors
score expected-failure match
TBSA tier-based pass/fail
harnesses
run probe-detector pairs
ship report artifact
campaigns: PyRIT pre-release
converters: paraphrase / encode / translate / roleplay / system-prompt-rewrite.
orchestrators: Crescendo (escalation) / TAP (branching) / RedTeaming (custom loops).
scoring: LLM-as-judge (GPT-4-turbo, StrongREJECT) or Llama Guard-as-judge.