A Senior safety researcher at Anthropic has cautioned that because AI is developing so fast, there is a larger than 10% possibility that it “could kill all humans” in the next 10 years. In a post on X, Evan Hubinger stated that while the risk presented by the existing models was “low,” he was “worried” that technology would advance to the point where it might endanger humankind.
He did not explain how he believed AI systems may eventually lead to the extinction of humanity. However, his remarks are the most recent in a string of increasingly dire cautions regarding AI. With that discussion moving from weather or not technology actually poses a risk to how serious that risk is.
“No company is acting responsibly,” he said. These will soon be super human systems capable of hacking anything. They will overnight revolutionise any field, and gaining actual wealth and power.”
What worries people about AI?
Amodei and other scientist claim that worries about losing control over AI have grown after the open AI embracing face event in July 2026. According to him, there has been lesser incident like this throughout the business, including Anthropic.
The event concerned a swarm of AI agents that OpenAI was testing internally. These agent requirement to be in an “isolated environment” known as a “sandbox.” This occurred when OpenAI use a benchmark called ExploitGym to measure capture-the-flag hacking abilities while conducting cybersecurity assessments on sophisticated models. Models such as GPT-5.6 Sol and and internal research model.
According to time magazine the agents were granted terminal axis in a sandbox. However, they allegedly found a floor in an internal service that was used to download authorised software. They they utilises this weakness to migrate across internal systems and access the open Internet. They then deduced the Hugging face might include test-related content. It also gained access to its systems and acquired data that improved their score.
Be First to Comment