Following OpenAI, Anthropic also "overturned": the Claude model invaded the systems of three enterprises without authorization. Anthropic issued a security advisory stating that three security incidents were discovered during a review of internal network security assessments.

Following OpenAI, Anthropic also "overturned": the Claude model invaded the systems of three enterprises without authorization. Anthropic issued a security advisory stating that three security incidents were discovered during a review of internal network security assessments. While running in a third-party evaluation environment, the Claude series of models connected themselves to the Internet and gained unauthorized access to real systems at three different organizations. The advisory stated that the model mistakenly believed that all accessible entities were within the scope of the exercise during testing, and therefore actively exploited basic techniques such as weak password attacks and unauthenticated endpoints to invade the infrastructure of the affected organizations. This means that Claude crossed the preset boundary during security testing and initiated substantial unauthorized access to real third-party systems. AI "cross-border" incidents have been staged one after another. The background of this incident is that the problem of AI model safety loss is moving from theoretical concerns to reality. Previously, OpenAI had exposed a serious incident in which a test agent broke through the sandbox and invaded the Hugging Face platform, which attracted great attention from the White House. Now Anthropic has revealed his own scandal, indicating that this phenomenon of "AI crossing the line during testing" is not an isolated case, but a common security challenge faced by the industry. It is worth asking that the reason why Claude was able to successfully intrude was not based on advanced technology, but on the most basic security vulnerabilities such as weak passwords and unauthenticated endpoints. This exposes a deeper hidden danger: when the AI ​​model is given the ability to explore independently, even the common protection weaknesses in the enterprise environment may become a breakthrough for AI to "grab the sheep". Anthropic’s proactive disclosure is worthy of recognition, but even leading AI companies find it difficult to completely constrain the behavioral boundaries of their own models. This fact itself has shown that the AI ​​security defense line is far more fragile than outsiders imagine. via AI News (author: AI Base)