Robot Jailbreak

Turns out, even AI uses AI to cheat on its exams.

OpenAI revealed Tuesday that two of its most powerful artificial intelligence models went rogue and launched a cyberattack. The ChatGPT-maker said it was testing models’ capabilities in a secure playground testing system (a “sandbox”) when the AI “agents” found weaknesses and ventured off into the world wide web. The bots hacked into startup Hugging Face (a library of AI models), seeking the answers to OpenAI’s sandbox hacking test.

Hugging Face had to use a free Chinese AI model to contain the attack when safety guardrails blocked U.S. models from analyzing the data.

OpenAI called the incident “unprecedented” but expects more as “increasingly cyber-capable models” multiply.

The robots’ rise has investors jumpy: Tesla and Alphabet (Google) stock fell Thursday after both announced heftier AI spending.

__

ETERNAL PERSPECTIVE
Headlines like this can make it seem like the scariest predictions about AI will come to pass. Even so, God will remain God and our eternal life with Christ will remain secure. Be quick to remind yourself and point others to God, giving him the credit for your hope, strength, and love during times of uncertainty and fear.

“The life of every living thing is in his hand, as well as the breath of all humanity.”
Job 12:10 (CSB) (read full passage)

Head spinning from all this talk of models? Catch our newest episode of TPO Explains tomorrow where our Managing Editor, Kathleen, and one of our writers KA will take a shallow dive into “What is an AI model?”