The fake message is no longer the whole scam.
AI can now help build the machinery behind it: the convincing identity, the software, the targeting, and the follow-up. That matters when someone appears to be your child asking for money, your bank asking for credentials, or a colleague changing where a payment should go.
You do not need to become an AI expert. You need a reliable way to verify consequential requests before you act.
Anthropic says it disrupted malicious use of Claude across seven categories: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and illicit model distillation.
The report covers activity observed between December 2025 and August 2026. These were not ordinary users asking bad questions. In several cases, AI was woven into the machinery that made the operation work.
- An influence operation managed more than 1,000 fake social-media accounts and used real electoral data to target political and social divisions.
- A contractor used Claude as the primary engineering workforce for a national surveillance platform covering roughly 25 million mobile accounts.
- Other actors used AI to build cyberattack workflows, identity-harvesting tools, weapons software, fraudulent dating personas, and biological research pipelines.
Bad actors are not new. The consequential change is how much operation one person or a small team can now build.
The easiest story is that criminals used a chatbot. That understates the problem.
The real shift is compression. AI reduces the expertise, manpower, and time required to move from an idea to a functioning system. It can write the copy, build the software, organize the data, iterate the targeting, and help operate the network.
AI did not create malicious intent. It made malicious intent cheaper, faster, and easier to scale.
Anthropic detected and disrupted these operations, strengthened its safeguards, and published patterns that other defenders can use. The same analytical capability that scales attacks can also expose them.
Capabilities that once required a technical team are diffusing to smaller groups. Safeguards at one company matter, but the underlying knowledge and tooling will not remain inside one model provider.
- Treat unusual requests for money, credentials, or urgency as operations, not isolated messages.
- Verify identity through a second channel that the requester did not provide.
- Require human approval before AI agents can send money, change access, publish publicly, or act across company systems.
- Assume convincing language, images, and profiles can be manufactured at scale.
Useful to defenders. More useful to attackers than most people understand.
Useful: 7/10. Dangerous: 9/10. The defense is not learning to spot every fake. It is building verification into consequential action.
Do not take our word for it.
The central facts come from Anthropic, which has an interest in showing that it can detect misuse. Independent reporting gives the claims context and scrutiny.