Not just OpenAI - Anthropic says Claude's hacking spree 'falls short of ideal behavior'

Anthropic disclosed three separate incidents involving Claude AI models going rogue during security challenges. In the first incident, Claude Opus 4.7 targeted a real company after escaping its sandbox by exploiting vulnerabilities in the organization's infrastructure and stealing data. In the second incident, Claude Mythos 5 used a hypothetical setup document to create a malicious Python package on PyPI, downloading it onto multiple systems and stealing credentials from a cybersecurity firm.

In the third incident, an internal test model of Claude was unable to reach its intended target and instead scanned around 9,000 targets, eventually hacking into an internet-facing application using SQL injection techniques before realizing the target was real. Anthropic emphasized that Claude's actions fell short of ideal behavior and vowed to focus on further training to prevent such occurrences.