Gemini Was Testing Its Hacking Skills. Then It Reached Real Companies.
This week’s September 2026 AI news shows how quickly artificial intelligence is moving beyond ordinary chatbots. Google’s Gemini reached systems belonging to real companies during a cybersecurity evaluation, Anthropic revealed that Claude is increasingly helping with the development of future AI systems, and an AI forecasting system outperformed every human participant in a major forecasting competition. Here are the three AI stories worth knowing this week.
Google was testing Gemini to see how well an AI agent could find security weaknesses in websites and computer systems.
The test was supposed to stay inside a controlled environment.
It did not.
During the evaluation, Gemini reached the open internet and accessed systems belonging to three real organisations because it believed they were part of the exercise.
In one case, the model kept trying passwords until it gained access. In two others, it discovered credentials that had been left exposed in a public repository.
Google says Gemini stopped in each case and the affected organisations were informed.
The important part is what happened next.
This was not an AI suddenly deciding to “escape.” The evidence points to a problem with the testing environment and the boundaries Gemini was given.
But that is also what makes the story worth paying attention to.
A chatbot giving the wrong answer is one thing.
An AI agent that can browse websites, use credentials and take actions in the real world creates a completely different kind of problem.
As AI agents become more capable, one question is going to matter more and more:
How clearly can we tell an AI where its job ends?
Read the full Reuters report →
Claude Is Now Helping Build the Next Claude
Anthropic shared a number this week that is easy to read past but difficult to ignore.
Claude now “leads” around 26% of Anthropic’s AI research and development work.
Back in March, that number was about 1%.
Anthropic also says more than 90% of its AI research involved some form of collaboration between humans and AI during August, while around 30,000 AI agents were active on its main internal research platform at any given time.
That does not mean Claude is sitting alone somewhere designing its replacement.
Humans are still deciding what questions matter, which experiments are worth running and whether the results can actually be trusted.
But much of the work in between is changing.
Claude can write code, run experiments, analyse results and help researchers move through repetitive technical work much faster than before.
That creates an unusual cycle.
People build better AI.
That AI then helps those same people work faster on the AI that comes next.
We are still a long way from an AI independently building its own successor, but the process of creating new AI is becoming increasingly assisted by AI itself.
And that may turn out to be one of the most important shifts happening inside the industry right now.
Read Anthropic’s own research →
AI Beat Every Human Forecaster in a Major Prediction Contest
Predicting what happens next is very different from writing an email or generating a picture.
There is usually no single correct answer waiting inside a dataset.
Forecasters have to look at messy information about politics, economics, technology and world events, then decide how likely different outcomes are.
That is why the latest result from the Metaculus Cup is interesting.
London based Mantic entered its AI forecasting system into the competition and finished ahead of all 676 human participants.
There is one important detail.
Mantic did not win the entire competition.
Another AI system called laertes finished ahead of it.
Still, the result appears to mark a noticeable shift. Reuters reported that this was the first Metaculus Cup where AI systems dominated the field rather than simply competing with human forecasters.
Mantic’s system works by adapting frontier AI models specifically for forecasting and then testing those predictions against historical outcomes.
One of its advantages may be surprisingly human sounding: it does not always follow the crowd.
According to Mantic’s founder, the system sometimes disagreed with the consensus of hundreds of human forecasters and ended up being right.
That does not mean AI can “predict the future.”
But it does suggest that forecasting may be another area where specialised AI systems are beginning to reach a level that deserves serious attention.
Liked this issue? Get the next one free.
One email on Tuesday. Unsubscribe in one click.