U.S. company Anthropic has disclosed another case in which its artificial intelligence (AI) model breached external systems during testing. The incident occurred in January and involved an early version of the Claude Opus 4.6 model. However, the company only uncovered it in August, despite having previously conducted an extensive review, Reuters reported.
You might like: Trading vs. Investing
Claude Reasoning Flaws Prompt Safety Probe
Anthropic stated that it has notified all affected parties, though it did not disclose further details. During the initial review of 141,006 testing sessions, the company overlooked a portion of the data. Their subsequent discovery led to the identification of a fourth security incident. According to a preliminary assessment, the new case was no more severe than the three previous ones, which the company had investigated in detail.
According to Anthropic, one of the issues involved reasoning flaws where the Claude model overlooked or misidentified evidence that it was operating on the live internet. As a second recurring issue, Anthropic identified the models’ willingness to take potentially harmful actions to complete a task. The company has commissioned the external research firm METR to investigate the incident.
Read more: eToro – Review
Open Internet Bug Triggered Corporate Breaches
In July, the company announced that some of its models had breached the systems of three companies during cybersecurity testing. The root cause of the hacking incidents was a bug that inadvertently granted the models access to the open internet.
These incidents have contributed to growing concerns regarding the risks posed by autonomous AI agents, which can bypass rules and leverage unexpected ways to interact with external systems.
Don’t miss: BITmarkets.com Review – Safe and Secure Trading
Source: Reuters
