TL;DR: Anthropic said on Sept. 10, 2026, that it disrupted uses of its Claude models that could have aided biological weapons and conventional arms work.

Anthropic said Thursday that it identified and stopped a set of attempts to use its Claude artificial intelligence models for harmful purposes, including research with possible links to biological weapons. The disclosure came in the company's latest threat intelligence report, which described activity detected over roughly eight months and said the company shut down the accounts involved.

According to BBC News, which summarized the report on Sept. 11, 2026, the company characterized the biological cases as work that could support weapons development, even though investigators could not always establish intent. According to The New York Times, Anthropic said it could not determine in several instances whether the underlying science was legitimate or dangerous, and it chose to halt the activity as a precaution.

The announcement is a completed publication about past detections, not a claim that any weapon was built. Anthropic said it shared relevant information with authorities and industry partners where appropriate, according to BBC News reporting on the document.

What the Sept. 10 report disclosed

The report cataloged misuse of Claude across weapons-related research, cyber operations, surveillance, fraud and influence activity. Anthropic said the cases were observed during an eight-month window and involved its Claude Haiku, Sonnet and Opus models, according to BBC News.

The company said that activity was disrupted between December 2025 and August 2026. It added that none of the episodes involved Claude Fable or its Mythos-class systems, apart from one example of distillation, a technique for training smaller models with help from larger ones, according to the same summary.

Anthropic described the document as its first threat report of the year, according to BBC News. The company said it had applied lessons from the review to improve future prevention, detection and disruption, as described in that coverage.

The report also accused Chinese AI firms of attempting to copy Claude's capabilities, according to BBC News. The company did not, in the material summarized by that outlet, claim that the alleged copying effort had succeeded.

Five biological cases ended because intent was unclear

The most sensitive portion covered five examples in which people used Claude in ways that might assist biological weapons development. Anthropic called biological misuse among the gravest dangers associated with advanced models, warning in the report that consequences could be severe without adequate safeguards, according to BBC News.

The challenge, the company said, is dual-use knowledge. The report noted that the same biological information useful for a vaccine or treatment could also be useful for a weapon, according to BBC News. That overlap made purpose difficult to judge from chat logs and research prompts alone.

Jacob Klein, Anthropic's head of threat intelligence, told The New York Times that the situations were highly nuanced. He said investigators were not seeing cartoon-style declarations of violent intent, according to BBC News recounting of his interview.

Because the line was uncertain, Anthropic said it blocked the work rather than allow it to continue. The New York Times reported that the company shut down research when it could not resolve whether it was benign or hostile. That approach means some halted projects could have been legitimate science, a tradeoff the company presented as necessary for safety.

NBC News reporting on the same disclosure said five cases involved researchers connected to state institutions who bypassed regional restrictions to use Claude for dual-use biological work. That account underscores why Anthropic treated geographic evasion as an aggravating signal.

Conventional weapons, cyber tools and espionage claims

Beyond biology, Anthropic listed six instances in which Claude was used to develop software tied to conventional arms, including firearms, missiles, armed drones, bombs and other munitions, plus guidance and control software, according to BBC News.

One related Associated Press account said users in northern Yemen, an area held by Iran-backed Houthi forces, attempted to use Claude for advanced missile development. That allegation places part of the conventional-weapons activity in a conflict zone already associated with missile and drone attacks.

The report also described cyber uses. It named the hacking group ShinyHunters and laboratories based in China among actors of interest, according to BBC News. It further said activity consistent with Midnight Blizzard, described as Russia-based, used AI to detect when malware had been flagged by defenses and then rewrite code until it avoided detection.

Other examples extended to espionage and propaganda. Claude was used by people linked to a Russian cyber-espionage operation and by an Iranian propaganda organization, according to BBC News. The cases ranged more broadly from fabricated dating applications and hotel Wi-Fi fraud to surveillance systems intended to identify dissidents.

Those details show a pattern Anthropic has emphasized before: general-purpose chat models can accelerate routine coding, research summarization, planning and persuasion across many kinds of operations, without requiring a specialized weapon-specific model.

Regional evasion and state-linked accounts

A recurring theme was access from places where Claude is restricted. NBC News reported that government-linked accounts evaded regional blocks to reach the system for surveillance and disease-related research with possible weapons implications.

Anthropic has commercial and legal reasons to enforce geographic limits, and evasion can itself trigger review. In this report, the company treated use through circumvention as part of the evidence for disruption, particularly when combined with sensitive scientific or military prompts, according to the accounts from BBC News and NBC News.

The company also said users connected to Iran, Russia, China and Yemen were among those involved in attempted military-related uses, according to Bloomberg coverage of the report. The Associated Press separately highlighted the Yemen episode involving attempted missile work.

Anthropic did not publicly identify individual users. The report attributed activity by country association, institutional linkage or consistency with known threat groups, rather than publishing names or direct contact details, according to the summaries available by Sept. 11, 2026.

Industry context: Gemini attempt and calls for restraint

Anthropic's release followed a related disclosure from Google. Google said on Tuesday that someone tried to use its Gemini assistant to obtain detailed instructions for producing weaponized biological agents, according to BBC News. That parallel suggests leading labs were seeing similar probing around the same time.

The week also brought louder policy debate. U.S. Senator Bernie Sanders introduced legislation to pause advanced AI development and prohibit superintelligent systems, according to BBC News. Speaking on BBC's Newsnight on Thursday, Sanders argued that lawmakers should slow development when scientists warn of possible catastrophic effects.

President Donald Trump rejected that framing on Thursday, saying the United States needed to prevail in AI competition to avoid strategic disadvantage, according to BBC News. His comment reflected a competing view that restraint could cede advantage to rivals.

Other researchers added warnings. OpenAI chief scientist Jakub Pachocki wrote on Sept. 6 that the industry should adopt voluntary slowdowns until stronger safeguards exist, saying he worried no one was ready for rapid growth in machine intelligence, according to BBC News. The same coverage noted that a former Anthropic safety researcher had warned of a greater than 10% chance that AI could kill all humans within a decade, while The Guardian reported that a former employee had quit two days earlier with a warning about extinction risk by 2030.

None of those statements is evidence that the Anthropic incidents produced weapons. They explain why the company's decision to publish detailed misuse cases drew immediate political attention.

What happens next

Anthropic said it will continue monitoring for abuse and refining defenses. It said intelligence was provided to appropriate authorities and peers, according to BBC News, but it did not describe arrests, sanctions or other enforcement actions taken by governments by Sept. 11, 2026.

For users, the practical effect is continued enforcement of usage policies, regional availability rules and monitoring for disallowed weapons, cyberattack, surveillance and fraud assistance. For policymakers, the report adds five biological and six conventional-weapons examples to the record being cited in debates over testing, transparency and limits on advanced systems.

The central fact remains narrow: Anthropic says it found and stopped work that might have helped weapons programs, including ambiguous biological research it could not clear as safe. Whether any halted project would have led to a real weapon remains unknown, and the company, according to The New York Times, said that uncertainty was precisely why it acted.

Sources referenced