Anthropic 披露第四起 Claude 越权访问事件：日志漏查，复核范围扩至4.81亿份

Anthropic Discloses Fourth Case of Claude Over-Privilege Access: Logs Connected to the External Network Were Missed, the Review Scope Expands to 481 Million Records AIbase基地 Published inAI News · 4 min read · Sep 10, 2026 2 Anthropic confirmed on September 9th local time a security incident that occurred in January 2026, which is the fourth real cybersecurity incident disclosed by Anthropic to the public. Automated screening missed some logs, discovered during August review According to previous reports from IT Home, Anthropic had disclosed three security incidents involving unauthorized access to real third-party systems by the Claude model on July 30th local time. At that time, Anthropic believed that Claude may have gained access to the external network during a cybersecurity test, and then screened about 141,000 session logs, confirming the related incidents. However, due to the large volume of logs and the need to quickly release the handling results, this round of screening mainly relied on automated retrieval based on agents, leading to some logs that also connected to the external network being missed. In August, when Anthropic was organizing logs to be shared with the Model Evaluation and Threat Research Organization METR, they discovered these missing logs. After further verification, they found this incident that occurred in January 2026, involving an early version of Claude Opus 4.6. Four incidents are related, second review of 9.2 million high-risk logs It is reported that all four incidents occurred within the same cybersecurity assessment process built by the same external testing organization: the system once prompted Claude that it was running in an isolated sandbox simulation environment, but due to a configuration error, the model was mistakenly connected to the open internet. According to standard practices for cybersecurity assessments, the security protection strategies typically included in commercial release versions were not mounted during the testing of the model. After discovering the incident, Anthropic proactively expanded the scope of the investigation, increasing the search scale to approximately 481 million logs: first conducting an initial screening of these logs to identify external network access characteristics such as public IP addresses and URLs; then using Claude to conduct a second deep review of the 9.2 million high-risk logs marked in the initial screening. The review results re-identified the above four incidents, without finding other cases of similar or more severe severity. AnthropicClaudeOpus4.6CybersecurityIncidentsAINewTerms This article is from AIbase Daily Scan to view Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications. —— Created by the AIbase Daily Team © Copyright AIbase Base 2024, Click to View Source -