Skip to content

Paper 038 / Anthropic / DeepMind

Red Teaming Language Models with Language Models

This paper uses language models to generate adversarial test questions for other language models, increasing the scale and diversity of red-team evaluation.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / alignment_and_safety

This paper uses language models to generate adversarial test questions for other language models, increasing the scale and diversity of red-team evaluation.

CONTRIBUTION / 01

Contribution 1

Automated red-team prompt generation with language models.

CONTRIBUTION / 02

Contribution 2

Compared zero-shot, few-shot, supervised, and reinforcement-learning approaches.

CONTRIBUTION / 03

Contribution 3

Created larger adversarial datasets and analyzed discovered failure patterns.

02

Research context

alignment_and_safety / 2022

Red Teaming Language Models with Language Models places red teaming inside the broader alignment and safety discussion at Anthropic / DeepMind, with adversarial evaluation supplying a second analytical lens. The editorial sequence connects three claims: Automated red-team prompt generation with language models; Compared zero-shot, few-shot, supervised, and reinforcement-learning approaches; and Created larger adversarial datasets and analyzed discovered failure patterns. The combination matters because safety only has meaning under the paper's stated setup. Operationally, the record points to one consequence: model-generated probes can expand safety testing, but automated attacks still require human review and coverage analysis.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Red Teaming Language Models with Language Models tests automated red-team prompt generation with language models and compared zero-shot, few-shot, supervised, and reinforcement-learning approaches against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

For Red Teaming Language Models with Language Models, the useful question is what the documented setup supports and where transfer remains untested. Its evidence chain begins with two documented moves: Automated red-team prompt generation with language models; and Compared zero-shot, few-shot, supervised, and reinforcement-learning approaches. Reported evidence then addresses: Created larger adversarial datasets and analyzed discovered failure patterns. The resulting interpretation is practical but conditional: model-generated probes can expand safety testing, but automated attacks still require human review and coverage analysis. Its boundary is that evidence for red teaming in Red Teaming Language Models with Language Models covers rater instructions, prompt sampling, selected threat model, evaluator models, model revisions, and construct validity; behavior beyond that documented envelope remains untested. Any extension should report how altered adversarial evaluation conditions affect the original red teaming result. A transfer experiment for Red Teaming Language Models with Language Models should preserve the red teaming reference, expose adversarial evaluation differences, and save evidence that narrows Compared zero-shot, few-shot, supervised, and reinforcement-learning approaches.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Red Teaming Language Models with Language Models supports created larger adversarial datasets and analyzed discovered failure patterns.
05

Practical implication for AI builders

Anthropic / DeepMind / 2022

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / red teaming

For a proposed BrokenGPT experiment based on Red Teaming Language Models with Language Models, run scheduled adversarial prompt generation against each endpoint, cluster new failures, and require human triage before adding cases to the regression suite. Keep the red teaming path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / adversarial evaluation

Validate the proposed red teaming route against the paper's reported outcome: Created larger adversarial datasets and analyzed discovered failure patterns. For Red Teaming Language Models with Language Models, record adversarial coverage, helpful-answer retention, and refusal precision; review adversarial evaluation failures separately before any proposed red teaming decision.

TRADEOFF / 03

Proposed decision boundary / safety

Balance usefulness, oversight burden, and residual risk before promoting the proposed safety design. Because the paper leaves deployment drift, unsampled behaviors, language coverage, adversarial adaptation, and judge bias as open implementation variables rather than consequences established by its experiments, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • Evidence for red teaming in Red Teaming Language Models with Language Models covers rater instructions, prompt sampling, selected threat model, evaluator models, model revisions, and construct validity; behavior beyond that documented envelope remains untested.
  • The paper leaves deployment drift, unsampled behaviors, language coverage, adversarial adaptation, and judge bias as open implementation variables rather than consequences established by its experiments.

PRIMARY SOURCES

  1. 01
    Red Teaming Language Models with Language Models

    Anthropic / DeepMind — Primary primary arXiv paper / 7 February 2022 / Ethan Perez, Saffron Huang, Francis Song, and 6 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Red Teaming Language Models with Language Models study?

This paper uses language models to generate adversarial test questions for other language models, increasing the scale and diversity of red-team evaluation.

02Which methods does Red Teaming Language Models with Language Models use?

The experimental design in Red Teaming Language Models with Language Models tests automated red-team prompt generation with language models and compared zero-shot, few-shot, supervised, and reinforcement-learning approaches against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Red Teaming Language Models with Language Models report?

The reported evidence in Red Teaming Language Models with Language Models supports created larger adversarial datasets and analyzed discovered failure patterns.

04What is the proposed BrokenGPT application for Red Teaming Language Models with Language Models?

Proposed: run scheduled adversarial prompt generation against each endpoint, cluster new failures, and require human triage before adding cases to the regression suite.

MAJOR LAB RESEARCH / PAPER 038

Continue after Red Teaming Language Models with Language Models

After Red Teaming Language Models with Language Models, browse the full index for adjacent alignment_and_safety research and work from Anthropic / DeepMind.

Open the paper index