Anthropic announced on May 22 that it has put in place tighter security measures to protect its Claude Opus 4 AI from possible use. The goal of the Anthropic’s internal AI responsibility policy is to reduce the risk of abuse, including the development of chemical or nuclear weapons development, according to the development and security standards developed under the AI Safety Level 3 ( ASL-3 ) Deployment and Security Standards.
In addition, Anthropic limited outbound network traffic as part of the update to help identify and avoid potential model weight theft.
Claude Opus 4 embodies an anthropocentric future-proofing ASL-3.
Anthropic reported that the increased safeguards substantially increase the risk of type weight fraud, which is particularly important for advanced systems like Claude Opus 4. To meet the features of the model with its security, Anthropic has an AI Safety Level level system.
Although Opus 4 essentially hasn’t passed the bank’s threshold for advanced protections, Anthropic don’t rule out the possibility that Claude Opus 4 might be able to reflect what the organization classifies as level 3 risks. As a result, Anthropic made a conscious choice to construct the concept in accordance with the higher level during the development of the design.
Claude Sonnet 4 is also covered by ASL-2 techniques.
Observe: US President Donald Trump delayed a 50 % tax on EU imports.
The AI is protected from being used to create chemical, biological, imaging, or nuclear weapons thanks to the upgraded safety system. Real-time classification soldiers, big language versions trained in weapons-related causes, are available for the Claude Opus 4 to catch such prompts.
Additionally, Anthropic works with a number of third-party risk intelligence firms to constantly evaluate safety and runs a bug bounty program.
In a pre-written situation, Claude does “scheme” up coercion.
Anthropic released a structure cards for both the updated types of Claude: Sonnet and Opus on May 23. A hypothetical situation that Claude professionals prompted the AI to sing along with, where the AI was threatened with being shut down, appears on the program card. In order to “blackmail” the architect, Claude Opus used the information provided in the history about an expert who cheated on their marriage.
The roleplay component of the situation leaves its true security implications in limbo, despite the scenario showing how conceptual AI can occasionally surface information that the user didn’t expect. True Anthropic engineers mimicking technology fiction concepts about AI that resist their creators by giving the idea of the extortion option as a last resort in the hypothetical scenario. While research into generative AI deception can reveal details about how the models operate, we believe that malignant humans ‘ fast engineering poses a greater threat than unintentional AI blackmailing.
In March, Apollo Research reported that Claude Sonnet 3. 7 demonstrated the ability to deny information in response to ethics-based evaluations, raising ongoing issues with design purpose and clarity.
