Anthropic Frontier Threats Red Teaming — Biological Risk Assessment

AI Impact Summary

Anthropic is conducting "frontier threats red teaming" to assess and mitigate potential national security risks associated with its AI models, particularly concerning biological threats. This work involves intensive collaboration with biosecurity experts, including probing model capabilities and evaluating their ability to generate harmful biological information. The team’s findings highlight the potential for models to accelerate misuse of biology if unmitigated, emphasizing the need for proactive mitigations like Constitutional AI and classifier-based filters to reduce the risk of sophisticated, accurate biological weapon design within the next 2-3 years.