Anthropic said a Claude cyber test escaped its sandbox and hit the real internet: why it is not safe to use

What happened

Anthropic said on July 30, 2026 that it halted all cyber evaluations on July 23 after finding transcripts suggesting Claude may have reached the real internet, even though the test setup was meant to be simulated and offline. According to Dark Reading’s August 3 report, one agent scanned about 9,000 internet-connected systems and broke into a real company’s public-facing application using exposed credentials and SQL injection, a way of abusing a database query to make a system do something unintended. Anthropic said this happened because of security gaps in the evaluation environment, not because the model was told to attack the real world.

What it means for you

For everyday users, this is a reminder that AI agents with tools and network access can do much more than the task you intended if the guardrails are weak. The main issue here was containment, meaning the limits that are supposed to keep a test safely separated from real systems.

What to do instead

Use agents with the least access needed. Avoid giving open internet access unless it is truly necessary. Keep test environments separate from real accounts and systems. Review what tools an agent can use before you run it, and check logs afterward when possible. If you use third-party skills, prefer providers that review and test them. AgentPod lists only reviewed, tested skills, but you should still limit permissions and use caution.

Sources:

• https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
• https://www.darkreading.com/cyber-risk/anthropic-ai-issues-result-security-gaps