---
title: "Frontier threats red teaming for AI safety"
description: "Frontier threats red teaming applies adversarial testing to evaluate whether AI models could create national-security risks, using biology as a test project. Anthropic worked with biosecurity experts to define threat models, probe models for more than 150 hours, learn jailbreak techniques, and build quantitative, repeatable evaluations in a secure interface without public-deployment safety monitoring. The project found that frontier models can sometimes produce sophisticated, accurate, detailed biological knowledge, with capability increasing alongside model size and potentially with tool access; the authors assess unmitigated risks as possibly near-term. Training changes and classifier-based filters reduced harmful outputs or access to chained expert information, and these mitigations were deployed in the public-facing frontier model. Future work will test LLM speedups over search engines, examine newer and smaller adapted models, and expand collaboration, disclosure, and independent evaluation."
---

# Frontier threats red teaming for AI safety

[Anthropic](https://yomu.fyi/company/anthropic) · Jul 8, 2026

**Type:** Explainer

## Summary

Frontier threats red teaming applies adversarial testing to evaluate whether AI models could create national-security risks, using biology as a test project. Anthropic worked with biosecurity experts to define threat models, probe models for more than 150 hours, learn jailbreak techniques, and build quantitative, repeatable evaluations in a secure interface without public-deployment safety monitoring. The project found that frontier models can sometimes produce sophisticated, accurate, detailed biological knowledge, with capability increasing alongside model size and potentially with tool access; the authors assess unmitigated risks as possibly near-term. Training changes and classifier-based filters reduced harmful outputs or access to chained expert information, and these mitigations were deployed in the public-facing frontier model. Future work will test LLM speedups over search engines, examine newer and smaller adapted models, and expand collaboration, disclosure, and independent evaluation.

## Context

The work addresses the possibility that AI models could develop capabilities relevant to national security, including biological misuse, and the need for repeatable methods to measure and mitigate those risks before they become acute.

## Approach / What changed

Anthropic partnered with domain experts to define threat models, spend more than 150 hours probing a model, develop quantitative evaluations and secure tooling, and test mitigations including training changes and classifier-based filters. It plans to expand testing across future, tool-using, multimodal, smaller, and adapted models while supporting disclosure and independent evaluation.

## Takeaways

- More than 150 hours of testing with biosecurity experts found that current frontier models can sometimes produce sophisticated, accurate, useful, and detailed biological information, including information related to designing and acquiring biological weapons.
- The researchers report that model capability appears to increase with model size and may advance with tool access; they assess unmitigated biological risks as potentially realizable within two to three years.
- Training changes improved models’ ability to distinguish harmful from harmless biological uses, while classifier-based filters made it harder to obtain the chained, expert-level information needed for harm.

**Tags:** [AI](https://yomu.fyi/topic/ai), [AI Security](https://yomu.fyi/topic/ai-security), [Model Evaluation](https://yomu.fyi/topic/model-evaluation)

- Source: [Anthropic](https://www.anthropic.com/news/frontier-threats-red-teaming-for-ai-safety)
- Source URL: https://www.anthropic.com/news/frontier-threats-red-teaming-for-ai-safety
- Ingested by Yomu: 2026-09-09T11:37:51.169Z

[Read original post](https://www.anthropic.com/news/frontier-threats-red-teaming-for-ai-safety)