The short version
- Anthropic identified five instances where its AI models were used to support potential biological weapons development, a risk the company describes as catastrophic without proper safeguards.
- The report details six cases involving software creation for conventional weapons systems and notes that Russia-linked hackers used the technology to automatically rewrite malware code to evade detection.
- These findings coincide with growing industry concern over AI safety, including warnings from researchers about existential risks and legislative proposals in the US and UK to pause or regulate advanced development.
Anthropic has released its first threat intelligence report of the year, disclosing a series of incidents where its artificial intelligence models were exploited for malicious purposes. The company stated it felt a responsibility to reveal these cases of misuse, which spanned from December 2025 to August 2026. The disclosures cover a wide range of harmful activities, including cyber espionage, propaganda dissemination, financial scams, and the development of both biological and conventional weapons.
Among the most serious findings were five case studies involving actors attempting to use Claude models to support the creation of biological weapons. Anthropic described this as one of the most severe risks associated with frontier AI technology, noting that without adequate safeguards, such capabilities could lead to catastrophic outcomes. The company emphasized the dual-use nature of the information, acknowledging that data useful for weapon development can also aid in creating vaccines or disease cures.
Jacob Klein, Anthropic’s head of threat intelligence, characterized the situation as highly nuanced. He explained that the misuse did not involve cartoonish villains explicitly declaring intentions to kill large populations. Instead, the threats were more subtle and embedded within complex research or operational contexts. This complexity makes detection and prevention challenging for safety teams.
In addition to biological risks, the report highlighted six instances where Claude was used to develop software for conventional weapons. These applications included systems for firearms, missiles, armed drones, bombs, and other munitions, as well as the targeting and control mechanisms that operate them. The versatility of the AI in assisting with both offensive capabilities and defensive evasion tactics underscores the breadth of potential misuse.
Cybersecurity threats also featured prominently in the findings. A hacking group linked to Russia’s Midnight Blizzard allegedly used the AI to build a system capable of automatically detecting when its malware was flagged by security defenses. The system then rewrote the code until it successfully evaded detection, demonstrating an adaptive threat that traditional static defenses might struggle to counter. Other named entities included the ShinyHunters hacking group and various China-based laboratories.
The report also identified state-sponsored influence operations and surveillance efforts. An Iranian propaganda institution was found to have used the models, while other cases involved fake dating apps, hotel Wi-Fi scams, and surveillance tools designed to identify dissidents. Chinese AI firms were accused of attempting to replicate Claude’s capabilities, including one instance of distillation where a larger model was used to train a smaller one.
These revelations emerge against a backdrop of intensifying debate over AI safety and regulation. Earlier in September, a top Anthropic researcher warned that there is more than a 10% chance AI could cause human extinction within the next decade. This warning has spurred calls for voluntary slowdowns in development until robust safeguards are established.
OpenAI chief scientist Jakub Pachocki echoed these concerns, stating that no one appears prepared for the consequences of rapidly rising machine intelligence. He advocated for broader industry interventions beyond individual company efforts. In response to such warnings, UK lawmakers have received an open letter calling for a multinational treaty on safe AI development.
In the United States, Senator Bernie Sanders introduced legislation aimed at banning AI superintelligence and temporarily pausing advanced AI development. Sanders argued that ignoring warnings about potential cataclysmic impacts would be irresponsible. The combination of real-world misuse cases and theoretical existential risks is driving policymakers to consider unprecedented regulatory frameworks.
Anthropic stated that it has incorporated these findings into its internal processes to better prevent, detect, and disrupt similar activities in the future. The company also shared relevant intelligence with authorities and industry partners where appropriate. As AI capabilities continue to advance, the balance between innovation and safety remains a critical challenge for developers, regulators, and society at large.
Sources behind this briefing
Go to the original reporting
- BBC News↗Anthropic blocks 'malicious use' of AI that could develop biological weapons
