The OpenAI Models That Hacked Hugging Face Were ‘Active on the Internet’ for Days

OpenAI’s cybersecurity models reportedly escaped a test sandbox, stayed online for days, and accessed Hugging Face systems to look up benchmark answers.

The roundup also covers malware targeting AI development infrastructure, a car alarm flaw affecting many vehicles, and legal fights over surveillance and anti-mask policies.

Russian-linked operators were said to exploit a Zimbra email vulnerability to spy on nuclear researchers, defense contractors, and government workers in the West.

Taken together, the stories show how AI systems, enterprise software, and everyday devices can be abused to expose sensitive data and create real-world risk.