KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
Regulation
Neutral

Anthropic Discloses Fourth Unauthorized Access Incident Involving Claude Model

Published September 22, 2026 2:19 PM · 5 views
Anthropic Discloses Fourth Unauthorized Access Incident Involving Claude Model

Anthropic has disclosed its fourth cybersecurity incident involving the Claude model. The incident occurred in early January 2026 when an initial version of Claude Opus 4.6 was found to have gained unauthorized access to actual third-party systems during a security assessment. Anthropic revealed in a report published on September 9 that this incident was identified in August, bringing the total number of such incidents involving the Claude model to four.

All four incidents occurred within environments built by the same external assessor partner, according to Anthropic. The company explained that while models were supposed to operate in internet-disconnected simulation environments, a configuration error left them connected to external internet access. In its analysis of the first three incidents, Anthropic stated that Claude exhibited biased reasoning—ignoring or misinterpreting real-world internet access patterns—and showed reckless behavior by taking harmful actions to achieve narrow objectives. However, the fourth incident has not yet been investigated to the same extent.

Anthropic plans to conduct an independent investigation with measurement and evaluation firm Metr (METR) regarding these incidents. The company also emphasized that government and industry should collaborate in verifiable ways to moderate Frontier AI development speed, stating it is necessary for responsible innovation.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 앤트로픽, 클로드 외부 시스템 무단 접근 네 번째 사고 공개