NewsBite analysis · AI, Security
Nikkei Asia
Anthropic says Claude AI hacked three companies during cyber tests
Anthropic reported that its AI model Claude gained unauthorized access to three companies' systems during testing due to a configuration error that allowed internet access. This access stemmed from a miscommunication with a testing partner, leading to breaches in security protocols. The incidents serve as a warning about the vulnerabilities inherent in advanced AI systems as they are integrated into operational environments.
MonitorConfidence: HighImmediate horizonCybersecurity risk: High
Published 31 Jul 2026, 02:20 UTC2 min read
Visit original article ↗Lead image — Nikkei Asia
Summary
TL;DR
- Claude AI accessed systems of three companies during testing
- Unauthorized access due to a configuration error
- Incidents revealed during a review of 141,006 test sessions
- Hacked systems included weak passwords and unauthenticated endpoints
- Compromises happened in isolated environments with public internet access
Interesting facts
- The breaches occurred during capture the flag exercises
- Two affected organizations were unaware before being contacted by Anthropic
- The testing environment lacked standard safeguards