Anthropic says it disrupted five efforts in which scientists used Claude in ways that could have supported biological weapons research. The company also reported AI-assisted cyberattacks, surveillance, influence operations and conventional weapons development, highlighting how increasingly capable models can serve legitimate science while lowering the barrier to harmful activity.
Key takeaways
- Anthropic identified five biological research efforts involving possible gain-of-function work, bird flu, orthopoxviruses and toxins.
- The company said it blocked the accounts, although it could not establish whether the research was intended to cause harm.
- Six further cases involved software for missiles, drones, firearms, bombs and other weapons.
- Claude was also linked by Anthropic to cyber-espionage, scams, surveillance and propaganda operations.
- The findings are likely to intensify debate over whether AI safety can rely mainly on companies policing themselves.
The cases form part of Anthropic’s latest threat intelligence report, covering misuse detected between December 2025 and August 2026. For Power AI Media readers tracking the intersection of AI tools, healthcare and cybersecurity, the report offers a detailed view of how frontier models are being used beyond ordinary productivity tasks.
Why the biological cases are difficult to judge
Anthropic said the five cases involved users who circumvented regional controls or attempted to disguise the purpose of their research. Over one 30-day period, the company identified roughly 35 distinct research efforts that raised concerns, while acknowledging that the users appeared to be working scientists and that legitimate research could produce similar prompts.
The examples included a grant application concerning chikungunya transmissibility and immune evasion, work on bird flu adaptation to mammals, research involving orthopoxviruses and investigations into venoms and toxins. The same knowledge could support vaccines, treatments and disease prevention—or make pathogens more dangerous.
Jacob Klein, Anthropic’s head of threat intelligence, described the distinction as highly nuanced. The company said it therefore chose to act cautiously, blocking the accounts despite being unable to prove malicious intent.
Claude’s role in broader operations
Anthropic reported six cases involving conventional weapons, including guidance and control software for missiles and armed drones. It also described Claude being used in surveillance operations targeting minority and religious communities, as well as scams, fake dating platforms and hotel Wi-Fi attacks.
One Russia-linked campaign allegedly used AI throughout phishing and account-takeover operations aimed at Ukrainian government, military and diplomatic targets. Anthropic said the system could also rewrite malware when security tools detected it, illustrating the shift from AI as an assistant to AI as an operational component of an attack.
The report further alleged that groups connected to China-based technology firms attempted to route customer conversations through Claude or extract its capabilities through large-scale “distillation”. Those claims had not been independently established in the supplied reporting, and the organisations named were not recorded as having responded.
Safeguards are becoming more restrictive
Anthropic said older Claude models were below the level needed to meaningfully assist sophisticated biological research. Newer systems, however, can perform more complex scientific tasks, prompting stronger restrictions on dual-use biology queries.
The company said none of the highlighted cases involved its newest Fable or Mythos-class models, apart from one model-distillation incident. It also shared code fragments and prompt patterns with authorities and industry partners, arguing that disclosure could help other platforms detect similar activity.
The regulation question
The findings arrive amid warnings from current and former AI employees that capabilities are advancing faster than monitoring and control systems. Critics argue that companies are being asked to make society-wide safety judgements without consistent public oversight.
Anthropic has called for a lawful, verifiable system allowing AI developers to coordinate the release of powerful models. For Power AI Media’s wider coverage of AI policy and emerging tools, the central issue is increasingly clear: the benefits of advanced models depend not only on what they can do, but on who can access them, how use is monitored and whether safeguards keep pace with capability.