Anthropic has published a new report on the malicious use of its models, and the picture is much broader than traditional chatbot abuse. According to the company, between December 2025 and August 2026, actors were detected attempting to use Claude for cyberattacks, surveillance, political propaganda, and biological research activities with dual use potential. Associated Press analyzed the document, the third report of its kind published by Anthropic since March 2025, highlighting a decisive shift: model capabilities have grown enough to compel the company to strengthen controls that were considered less necessary on earlier systems.
Anthropic points out that the published cases do not represent the typical use of its products, but rather the most relevant and novel examples of identified hostile activity. It is an important distinction: the report does not demonstrate that the majority of users rely on AI for dangerous purposes. It does show, however, how motivated actors are experimenting with models to lower costs, speed up operations, and bridge technical gaps.
The biological case shifts the threshold the most
Among the incidents described, one involves a request for assistance in drafting a funding proposal tied to gain-of-function research on the chikungunya virus, focusing on transmissibility and immune evasion. Dual use research can have legitimate purposes, such as better understanding a pathogen or developing countermeasures, but that same knowledge can enhance the ability to make it more dangerous.
Anthropic states that the request was blocked. The most significant takeaway from the report, however, is the admission that newer models require stronger safeguards precisely because they are more capable. The company maintains that earlier systems such as Claude Opus 4 and Sonnet 4.5 were below the threshold at which they could substantially assist a sophisticated user in conducting dangerous biological research; for current models, it says, it cannot offer that same assurance.
Safety shifts when the model becomes competent
A filter that works on a relatively weak system may not be enough as the quality of responses improves. If the model has a deeper understanding of biology, programming, or chemistry, even a minor policy loophole can lead to far greater consequences. It is an intuitive yet often overlooked principle: safety is not a static feature of a product; it must grow alongside its capabilities.
This is why Anthropic claims to have introduced broader restrictions on dual-use biological prompts in its latest models, including Claude Fable 5. The decision curtails certain legitimate use cases that are hard to distinguish from dangerous ones, creating an unavoidable trade-off between openness and prevention.
Cyberattacks are becoming accessible to smaller groups
The report also details cyber activities in which AI is used to automate or accelerate parts of an operation. This does not mean a model can entirely replace a team of skilled hackers, but it can lower the barrier of expertise required for certain stages: writing code, analyzing a configuration, generating message variants, rapidly translating instructions, or exploring attack surfaces.
Anthropic notes that sophisticated operations can become feasible even for individuals or small groups in ways that would have been far more difficult just a year earlier. It is an inverted democratization effect: the same technology that makes programming easier for a beginner also makes it easier to automate hostile behavior.
Spyware and surveillance highlight the commercial side of abuse
Spyware vendors are also among the actors cited. This detail is significant because it moves the issue beyond traditional cybercrime. A commercial product can use AI to analyze data, build tools, or optimize surveillance activities, fostering a market where model capabilities are embedded into services sold to governments or private clients.
For AI labs, establishing controls based solely on user identity is therefore becoming difficult. A formally legitimate company can use tools in controversial contexts. What is needed instead are evaluations based on behavior, request patterns, and intended use.
Propaganda is already industrial
Anthropic claims to have observed groups capable of creating hundreds of social accounts that seemingly belong to ordinary people and coordinating content aligned in the same political direction. The operations described span various regions, from Russia and Iran to Turkey, the Gulf, South Asia, Africa, and Europe.
The advantage of AI is not necessarily creating propaganda that is more persuasive than human-made content. It lies in producing large volumes of it, adapting it linguistically, and managing synthetic identities at a lower cost. This makes attribution more difficult and increases the volume of noise that platforms must analyze.
Labs see the operation before social platforms do
An interesting passage in the report concerns the position of model developers. A social network sees a campaign when the content is published. An AI lab can observe part of it during preparation, when the threat actor prompts the model to generate profiles, messages, or strategies.
This position creates a new responsibility, but also a new defensive capability. Sharing indicators among labs, platforms, and authorities could make it possible to disrupt an operation before it reaches the public. At the same time, it raises privacy and surveillance concerns: monitoring abuse must not turn into the indiscriminate scrutiny of legitimate activity.
Anthropic cannot be the sole arbiter
The company maintains that it publishes these cases to help competitors, governments, and civil society recognize similar patterns. It is a useful step, but it leaves open a governance question: a private company deciding which requests are acceptable and which are too dangerous. John Thickstun of Cornell University, cited by AP, noted that this is an uncomfortable position because it involves societal-scale value judgments without a genuine democratic process.
That is why public standards and external oversight are becoming more important. Not every technical detail can be written into law, but principles covering incident reporting, capability evaluations, and access to high-risk sectors can be established outside corporate boardrooms.
The report is also a snapshot of the near future
The most serious takeaway is not that Claude has become inherently “dangerous.” It is that increasing capabilities continually alter what a malicious actor can achieve. A safeguard that is sufficient in 2025 may prove weak in 2026. Security must therefore be a continuous process of measurement and updates.
Anthropic claims to have thwarted the described incidents and used the experience to strengthen its systems. That is the positive side of the story. The other is that adversaries will keep trying. The report highlights the most realistic dynamic in AI security: not a single decisive battle, but an ongoing race between the capacity for abuse and the ability to detect it.
The more useful models become for science, coding, and research, the harder it becomes to build guardrails that block only harmful behavior without hindering legitimate use. It is a technical challenge, but not exclusively so. It requires a collective definition of which risks we are willing to accept and which controls we consider proportionate. Anthropic's third report makes it evident that this debate is no longer about a hypothetical future: the abuses are already here, and they are evolving alongside the models.
Sources