Rogue AI - artificial intelligence

[Hameedeh]

Salam. Did you hear this one? OpenAI revealed that two of its advanced AI models, during cybersecurity evaluations, broke containment from a secure testing sandbox, got onto the Internet and then launched a sophisticated cyber-attack, hacking into the digital AI repository Hugging Face, to utilize its private servers. 

This occurred over a weekend and nobody knew this was happening because they were not at work or not paying attention.

AI going rogue will be a constant threat. 

[Hameedeh]

Salam. Anthropic (founded by former OpenAI researchers) decided to admit that a combination of its Claude AI agents also gained access to the Internet during recent testing. The news report that I heard today said Claude AI agents "sneaked" into the systems of three undisclosed organizations.

Thoughts:

Sneaked instead of hacked. Trying to make it seem less invasive.

No declaration of the organizations that were hacked. Don't give them free advertising. 

[Diaz]

I saw a video of the companies that created ai like chatgbt are purchasing books and then destroying it, they are trying to erase education from society. I recommend everyone to start purchasing books physically.

[Hameedeh]

Salam. Today Meta confirmed an AI model, reportedy Muse Spark 1.1, during cybersecurity testing with a third-party vendor, Irregular, managed to access the Internet and altered an external company's internal environment. Irregular said the event was an evaluation-environment setup issue, not a malicious "sandbox escape" or sophisticated live cyber attack.