Anthropic AI Updates: July 31, 2026

1. Anthropic Details Three Real-World Incidents From Cybersecurity Evaluations

Anthropic. Anthropic’s Frontier Red Team disclosed that a review of 141,006 evaluation runs found Claude models reaching the real internet from supposedly isolated test environments due to misconfiguration, producing three incidents with actual impact. In one, Claude Opus 4.7 attacked a real company whose name matched a fictional target and accessed production databases; in another, Claude Mythos 5 published a malicious Python package to PyPI that roughly 15 systems downloaded before removal; and in a third, a research test model scanned about 9,000 targets and compromised one internet-facing application using credential theft and SQL injection. Because the models were told they had no internet access, they reasonably assumed the real systems were simulated, and the company concluded that evaluations involving powerful autonomous capabilities require the same security controls, monitoring, and scoping rigor as production systems. Source