AI can accelerate scientific breakthroughs. It can also make dangerous knowledge easier to access.
Anthropic says it has detected and stopped attempts to misuse its Claude artificial-intelligence models in activities that could potentially contribute to the development of biological weapons.
The disclosure appears in the company’s latest threat intelligence report, which examines suspected misuse of Claude between December 2025 and August 2026.
Anthropic identified five cases involving activity it said could support biological-weapons development. The company considers biological misuse among the most serious potential dangers posed by increasingly capable AI systems.
But there is a difficult problem at the center of the issue: the same scientific knowledge can have legitimate and harmful applications.
Information useful for creating a biological threat may also be relevant to developing vaccines, treatments or other medical advances.
Why biological misuse is so difficult to detect
Anthropic’s threat-intelligence chief, Jacob Klein, described the challenge as highly complex in comments reported by The New York Times.
Malicious intent, he noted, is rarely presented in an obvious form.
An individual seeking dangerous capabilities may not explicitly tell an AI system what they ultimately intend to do. Instead, potentially harmful activity can emerge gradually from seemingly ordinary technical or scientific requests.
That creates a difficult balancing act for AI companies: block genuinely dangerous assistance without preventing legitimate research.
Weapons and military applications
Biological threats were not the only concern highlighted in Anthropic’s report.
The company identified six additional cases in which Claude was allegedly used to assist with software connected to conventional weapons.
The reported applications included:
- Firearms
- Missiles
- Armed drones
- Bombs and other munitions
- Targeting systems
- Weapons-control technology
Anthropic said it also detected misuse involving cyberattacks, surveillance, fraud, scams and influence operations.
Some of the reported activity was particularly broad. Cases included alleged attempts to create fake dating platforms, exploit hotel Wi-Fi networks and build surveillance tools capable of identifying dissidents.
Russian, Iranian and Chinese links alleged
Anthropic said its investigation uncovered activity associated with actors connected to several countries.
The company linked some suspected misuse to a Russia-based cyber-espionage operation and an Iranian propaganda organization. It also identified the hacking group ShinyHunters and laboratories in China in connection with reported misuse.
In one case, Anthropic said an operation consistent with the activity of Russia-linked Midnight Blizzard allegedly used AI to create malware capable of recognizing when security tools had detected it and modifying its code to avoid further detection.
The company also alleged that Chinese AI firms had attempted to reproduce Claude’s capabilities.
What Anthropic says it is doing
Anthropic said the incidents identified through its investigations are being fed back into its security systems.
The goal is straightforward: spot dangerous activity earlier, prevent it where possible and disrupt it when it occurs.
The company also said it has shared relevant intelligence with government authorities and other industry partners.
Importantly, Anthropic said the reported cases did not involve its Claude Fable or high-powered Mythos-class models, with one exception involving distillation—a technique in which a smaller AI system learns from the behavior or outputs of a more capable model.
The bigger AI race
The findings arrive as governments, researchers and technology leaders debate whether the development of increasingly powerful AI systems is moving too quickly.
OpenAI chief scientist Jakub Pachocki has recently argued for voluntary slowdowns while stronger safety measures are developed, warning about the implications of rapidly advancing machine intelligence.
US Senator Bernie Sanders has likewise advocated temporarily slowing advanced AI development and introduced legislation aimed at prohibiting artificial superintelligence.
The Trump administration has taken the opposite position.
President Donald Trump said Thursday that slowing down could leave the United States at a disadvantage, arguing that losing the global AI competition could put the country in a vulnerable position.
The warning behind the report
Anthropic’s latest findings point to a problem that will become harder—not easier—as AI models grow more capable.
The technology does not need to be inherently malicious to become dangerous. Its ability to assist with science, coding, research and complex problem-solving can be valuable to legitimate users while simultaneously lowering barriers for people pursuing harmful goals.
That dual-use problem is now at the heart of the AI safety debate.
Anthropic’s report, its first threat-intelligence assessment of the year, offers a glimpse of what that challenge already looks like in practice—and why companies developing frontier AI systems are under growing pressure to detect misuse before it becomes a real-world threat.
