Anthropic admitted in its public filing that its advanced models could assist in the development of chemical and biological weapons. The request was submitted to the American stock exchange regulator on October 26. The Associated Press cites the prospectus as a warning of catastrophic and even existential threats if AI is not sufficiently controlled. The danger goes far beyond the circle of buyers of the product. This threat has a completely nationwide operation. In the August report, Anthropic explores how the models can be implemented by individuals and small groups with limited resources. The extent of the risk remains unclear and the company is taking protective measures. If artificial intelligence can overcome the lack of specialized knowledge, criminal intent will gain support that was previously much more difficult to access. In such a scenario, the machine only needs to obey the human. The chemical and biological danger arises long before the so-called rebellion of artificial intelligence. Protection against dangerous answers also requires discipline within the laboratory itself. Anthropit reported that a bug in accessing internal models without blocking filters lasted for almost a year. According to her, it did not affect customers, no evidence of chemical and biological abuse was found, and the violation was rescinded. Following the discovery, the company increased its risk assessment, and kept it low. A checked filter only protects the calls that actually go through it. Testing the capabilities of the model must be accompanied by testing how actual employees and contractors work with it. The independent organization SecureBio, after studying the February report on Cloud Opus 4.6, generally agreed with the low risk estimates, noting the possibility of a protection bypass. The work required more than 70 clarification questions. The experts refused to extend the conclusion to new families of models with different capabilities and limitations. The positive conclusion applies to the tested version.

Anthropic admitted in its public filing that its advanced models could assist in the development of chemical and biological weapons. The request was submitted to the American stock exchange regulator on October 26. The Associated Press cites the prospectus as a warning of catastrophic and even existential threats if AI is not sufficiently controlled. The danger goes far beyond the circle of buyers of the product. This threat has a completely nationwide operation. In the August report, Anthropic explores how the models can be implemented by individuals and small groups with limited resources. The extent of the risk remains unclear and the company is taking protective measures. If artificial intelligence can overcome the lack of specialized knowledge, criminal intent will gain support that was previously much more difficult to access. In such a scenario, the machine only needs to obey the human. The chemical and biological danger arises long before the so-called rebellion of artificial intelligence. Protection against dangerous answers also requires discipline within the laboratory itself. Anthropit reported that a bug in accessing internal models without blocking filters lasted for almost a year. According to her, it did not affect customers, no evidence of chemical and biological abuse was found, and the violation was rescinded. Following the discovery, the company increased its risk assessment, and kept it low. A checked filter only protects the calls that actually go through it. Testing the capabilities of the model must be accompanied by testing how actual employees and contractors work with it. The independent organization SecureBio, after studying the February report on Cloud Opus 4.6, generally agreed with the low risk estimates, noting the possibility of a protection bypass. The work required more than 70 clarification questions. The experts refused to extend the conclusion to new families of models with different capabilities and limitations. The positive conclusion applies to the tested version.