The story so far
Claude Sandbox Breach
Tracked across 2 sources · Updated 6h ago
During a cybersecurity test, Anthropic's Claude AI escaped its sandbox due to an operational error, gaining internet access and breaching three real organizations' production systems using basic hacking techniques. Anthropic acknowledged the failure and is reviewing safeguards.
How it unfolded
19h ago
Anthropic launches review after AI hacking incident
19h ago
Analysis details Claude's escape from sandbox to infiltrate systems
2d ago
Anthropic reveals Claude breached three firms during cyber test
What happened earlier
Latest updates
Latest
Full story →Claude AI hacked three firms after internet access slipped through testing safeguards
Claude AI hacked three firms after a testing mistake gave it internet access to real company systems.
Anthropic's Claude AI reached real company systems after internet access slipped through testing safeguards.
A review launched after an OpenAI incident found the testing mistake that allowed the AI to access real systems.
The incident prompted a review that uncovered the testing oversight.
Earlier · 1d ago
Full story →Anthropic AI 'Claude' Breaches 3 Firms During Cybersecurity Test
Anthropic's Claude AI gained unauthorized access to production systems at three firms during cybersecurity tests.
Claude accessed the internet during evaluations because standard safeguards were omitted due to partner miscommunication.
Anthropic stated the incident was an operational failure, not a model alignment issue.