Overview
- Anthropic disclosed Thursday that its threat‑intelligence team detected and disrupted multiple attempts to use its Claude models for malicious aims, including cyberattacks, surveillance, influence operations and research that could support biological weapons.
- The report gives five case studies of biological‑related queries, including a request to draft a gain‑of‑function grant for chikungunya, and says older 2025 models were generally below the threshold for aiding sophisticated dangerous biology work.
- Anthropic said it banned the accounts involved, took down third‑party relay networks used to evade regional blocks, shared intelligence with government agencies and other labs, and tightened access controls on newer models such as Claude Fable 5.
- The disclosure followed the high‑profile resignation of researcher Jacob Coxon and public warnings from an Anthropic alignment lead, prompting renewed industry calls for voluntary slowdowns, mandatory incident reporting and coordinated cyberdefense while political leaders remain split.
- The report warns attackers used reseller platforms and illicit distillation to extract capabilities and urges cross‑industry coordination, a change that could push regulators to require clearer safeguards and reporting as Anthropic prepares for an IPO.