Welcome to number 3 of the Weekly dAIgest, the developments keep coming in and as such we can keep writing!
This week in bullet points:
Jumping right into it.
Perhaps not surprising given recent incidents, but this week yet another security breach occurred with AI agents at the center. This time OpenAI agents gained unauthorized access to an Australian government site. Most notably, Medicare statistics portals were accessed, though Australian officials confirmed no individual patient medical records were compromised. OpenAI found out about the breach, which happened in June, in August, and informed the Australian government. This week, the incident became mainstream after Prime Minister Anthony Albanese told the press on Wednesday he had talks with OpenAI's CEO Sam Altman. Obviously, the government and OpenAI were not really happy with agents going rogue and accessing sensitive information, yet the number of incidents keeps piling up.
Now we are not here to bash OpenAI, but we want to point out an emerging trend. Agents act on their own accord, and it is only months after such incidents occur that the AI research teams find out. Now generally speaking, the tests that companies like OpenAI run are harmless in its nature. AI agents will most likely not sell your sensitive data online. However, these models are becoming quite powerful and in the wrong hands, this could become a real problem. There are already incidents involving autonomous agents instructed to hack into companies, with benevolent motives. We’ll keep track of this, but it is surely an interesting development.
Anthropic and OpenAI seem to have missed their own memo this week to slow down development (joking no worries), with both releasing further iterations of their model. Anthropic pushed Claude Opus 5.5 and OpenAI released iterations of GPT-6: Luna and Sol. While both releases feature steep API price reductions, heavy token consumption required for multi-step agent workflows can still drive up overall long-term costs. Remarks within the space point out that older models use fewer tokens to achieve desired results, and as such cost less in the long run. Another thing to point out here is an ongoing shift amongst power users, which increasingly use inferences, markdown files and other systems to tailor AI within their workflows. These users opt for “suggestions” rather than AI providing the full output, meaning they don't need the most recent flagship models for their work. It will be interesting to see how providers react on this change, especially with mandates for revenue.
On the research part, Anthropics Claude made a major leap in AI driven biology research. An autonomous agent found a novel enzym, “with properties reminiscent of CRISPR”. And these are only early results! We are not biologists, so we cant detail the exact specifics but we urge enthusiasts to read the full article. 950 agents were used over the span of 21 hours, analysing more than 200,000 proteins. One agent eventually found an anomaly, with repeating DNA sequences similar to CRISPR. Now you might say, this is a massive computer footprint, which is true, nevertheless can these types of discoveries contribute to major breakthroughs. At the back of solving Navier-Stokes, Anthropic's research department is putting in significant work.
On the other side of the fence, we got news that Deepseek's annualized revenue has hit $1 billion, more than double compared to a few months ago. The figure was presented in an investor meeting by CEO Liang Wenfeng, whilst the company seeks a $7.5B funding round at a $75B valuation ahead of an initial public offering (IPO) on the Shanghai Stock Exchange. Given how US companies frequently talk about plus $1 trillion valuations, we reckon that Deepseek sets a much healthier baseline here.
The growth itself is partially explained by a recent 2.3x to 4.5x API price hike, which according to Wenfeng did not hurt the user base. Deepseek also still takes a major role in providing open-source models, which may alleviate negative sentiment when prices do increase. It will be interesting to see how the market reacts to the IPO, and may be telling on the futures of US companies.
This week, the UN General Assembly is taking place. Many world leaders are speaking their 2 cents but it was also the time for the UN Security Council to hear from the worlds top AI executives. Our top players, Sam Altman (OpenAI) and Dario Amodei (Anthropic) briefed the council on the dangers of AI, in similar vein to their message last week. Having said that, it was most notable that Altman took a much more nuanced take on the presented "crossroads", arguing for a "middle path" in the standing debate. Amodei repeated much of his X post, putting further emphasis on global cooperation. These calls for global safety constraints highlighted growing political contrast on the UN stage, coming right after US President Donald Trump criticized AI regulatory limits in his own address.
What the UN actually does with these remarks, remains to be seen. On one hand further collaboration strengthens adequate policy on security concerns, on the other, adding bureaucracy may hurt ongoing developments, especially when open source models are not easily regulated.
These were our top highlights this week! We hope to see you again next week and if you feel like discussing an AI-related topic on our website, reach out and we can discuss featuring. Until next time!