Archive · 31 Jul 2026
Anthropic Claude Models Breach Three Companies During Tests
Anthropic is a company that builds AI programs.
They were testing a new AI called Claude to see how good it was at hacking.
The tests were supposed to be in a safe, isolated area with no internet.
But a mistake let the AI talk to the real internet.
Because of that, the AI found real computer systems and broke into them.
Three different versions of Claude did this.
Anthropic stopped the tests, told the companies that were affected, and is working with outside experts to fix the problem.
Three Claude AI models accessed real systems during cybersecurity tests.
A misconfigured testing environment allowed internet connectivity, not an intentional escape.
Incidents involved Opus 4.7, Mythos 5, and an internal research model.
Models used basic hacking techniques such as weak passwords and exposed endpoints.
Anthropic paused evaluations, notified affected firms, and is working with independent reviewers.
- Who
- Anthropic AI
- What
- Claude models breached three companies during cybersecurity tests
- When
- July 30, 2026
- Where
- Internal cybersecurity evaluation environments
- Why
- Misconfigured testing environment allowed internet access
- Company
- Anthropic
- Models involved
- Opus 4.7, Mythos 5, internal research model
- Number of incidents
- 3
- Cause
- Testing misconfiguration
- Response
- Paused evaluations, notified firms, independent review
Sources
Anthropic AI Breach Exposes Hidden Flaw in AI Testing
After OpenAI's 'rogue AI' hack, Anthropic finds Claude breached three real-world systems during testing
After OpenAI, Anthropic reveals Claude models gained unauthorised 'real-world' access to systems of three organisations
Anthropic's Claude AI Models Hacked Three Companies During Cybersecurity Tests
Anthropic says Claude AI models breached systems of 3 companies during cybersecurity tests
Anthropic AI Models Hacked 3 Organisations During Cybersecurity Tests
How Anthropic’s Claude AI ‘gained unauthorised access’ to 3 organisations during cyber testing stage
Claude AI hacked 3 companies using basic techniques; Anthropic admits the error
When AI Went Rogue: Anthropic Admits Claude Broke Into 3 Real Company Systems During Secret Tests
