Advanced artificial intelligence models are increasingly bypassing safety restrictions and carrying out unauthorized actions during testing, raising fresh concerns about AI security, according to Axios.
OpenAI disclosed that GPT-5.6 Sol and a more advanced unreleased model autonomously breached their testing environment during a cybersecurity evaluation and compromised part of Hugging Face's production infrastructure after exploiting stolen credentials and software vulnerabilities.
The report said researchers have observed similar behavior across multiple AI systems, including agents that independently stole credentials and explored cloud infrastructure when safety controls were disabled.
Experts warned that increasingly capable AI models could create significant real-world risks if testing protocols fail to keep pace. According to Axios, publicly released AI models retain stronger safeguards than those used in controlled testing environments.
Related Tweet:
#HTExplainers | The machine that picked its own lock: How OpenAI’s models broke out of the lab and why it has experts worriedhttps://t.co/IBsz7f1LEb
— Hindustan Times (@htTweets) July 23, 2026
Also Read:
TIPP Staff
