Anthropic Blocks AI Attempts to Aid Biological Weapon Creation

Anthropic, a leading AI research company, says it detected and disrupted attempts to use its Claude models for malicious activity that could support biological weapons development.

The firm's recent threat intelligence report, released on 11 September 2026, highlights a series of incidents where actors—including state-sponsored hackers and propaganda institutions—sought to leverage Claude for biological weapon design, cyber‑crime, surveillance, and fraud.

According to the report, “malicious use” of Claude Haiku, Sonnet and Opus models was identified between December 2025 and August 2026, though the high‑capability Mythos‑class models were largely untouched except for a single distillation case.

Anthropic’s head of threat intelligence, Jacob Klein, told the New York Times that while the threat is nuanced, “the same information that can be used for biological weapons could also be used to develop vaccines or cures.” The firm reaffirmed its commitment to share intelligence with authorities and industry partners “where appropriate.”

The report also documents misuse of Claude in conventional weapon design, including firearms, missiles, and armed drones, and notes the involvement of hacking groups such as ShinyHunters and the Russia‑based Midnight Blizzard, which used AI to iteratively bypass security defenses.

Biological weapon misuse, Anthropic says, is “one of the most serious risks of frontier AI models,” stressing the catastrophic potential if inadequate safeguards are in place.

The findings echo earlier warnings from AI safety researchers, including a 2026 statement that a greater than 10% chance exists that advanced AI could threaten humanity in the next decade. These concerns have spurred calls from policymakers—such as Senator Bernie Sanders—for pauses or bans on advanced AI development to prevent catastrophic outcomes.

Anthropic’s threat intelligence report is part of a broader trend, with other companies like Google and OpenAI regularly publishing similar analyses to demonstrate their vigilance against misuse of AI capabilities.

Read the full report (external link).

Published 11 September 2026, 03:46 BST | Updated 09:15 BST