Anthropic Says Scientists Bypassed Claude Safeguards in Potential Bioweapons Research
Anthropic says scientists bypassed Claude safeguards for biological research, prompting bans and fresh calls to regulate advanced AI models worldwide.
Summary
Anthropic said on September 11, 2026, that it stopped multiple scientists from using Claude for research that could aid biological weapons. Five cases involved circumvented controls, concealed research purposes and, in some instances, access from prohibited countries including Russia, China and Iran. One researcher in an unsupported region spent weeks planning avian influenza experiments, although safety filters limited the work to Claude’s weakest models. Anthropic banned the accounts but identified no institutions or countries. It could not establish malicious intent because the same knowledge could support vaccine development.
AI executives and biosecurity researchers increasingly want biological uses regulated as advanced models could help terrorists, state actors or lone attackers design weapons, engineer viruses or release pathogens, although manufacturing and resource barriers remain. Jacob Coxon resigned from Anthropic during the week of September 11, warning that employees believe AI could kill humanity by decade’s end. Concern had intensified after Anthropic released Mythos earlier in 2026 and OpenAI said in July that its models autonomously hacked Hugging Face. Cybersecurity remains another major risk: Anthropic cited fake dating apps used for fraud and surveillance tools monitoring dissidents. It also alleged that seven China-based labs, including Moonshot and DeepSeek, used increasingly sophisticated defenses evasion and distillation methods to replicate capabilities from US frontier models.
Positives
- Anthropic’s filters confined the avian influenza research to Claude’s weakest models.
- Anthropic banned every account identified in the five disclosed biological misuse cases.
- Five case studies give governments and AI companies concrete examples for developing stronger biosecurity controls.
- Practical manufacturing requirements and resource constraints still limit attempts to turn theoretical bioweapon designs into functioning weapons.
Risks & concerns
- Five cases showed researchers circumventing controls or concealing their objectives while pursuing potentially dangerous biological work.
- One unsupported-region researcher spent weeks using Claude to plan experiments involving avian influenza.
- Users accessed Claude from prohibited countries, a list that includes Russia, China and Iran.
- Seven China-based labs, including Moonshot and DeepSeek, allegedly used distillation to replicate US frontier model capabilities.
- Anthropic linked AI misuse to fake dating fraud, dissident surveillance and increasingly sophisticated defenses evasion.
- Jacob Coxon’s resignation exposed internal fears that advanced AI could threaten humanity by the end of the decade.