At the VivaTech conference in Paris, discussions around AI clearly moved from technology to more realistic issues. Cyber-security risks, European concerns about reliance on external models, and corporate queries about returns on AI inputs became the three main lines of discussion.
Cybersecurity is the focus.
After the introduction of Mythos, OpenAI and GPT-55 Cyber, businesses and government concerns about the ability of AI to attack the Internet have risen. According to the article, such models are already able to detect unknown loopholes in key software more quickly and generate available attack codes.
This means that work that in the past required professional staff and a longer period of time is being implemented at a much faster and larger scale by the model. Markets are concerned that attackers may acquire such capabilities faster than most institutions repair systems and upgrade defence tools.
Peter DeSantis, Senior Vice President of the Amazon, said that AI might be better for the defensive side in the long run, but that short-term risks were more pronounced. A number of security teams are still adapting to new types of attack, and the speed of the defence system is not necessarily up to speed with model capabilities.
OpenAI has expanded the Daybreak project this week and combines GPT-5/Cyber with the Patch the Planet open-source patch project, which includes Cloudflare, Cisco and CrowdStrike, with the aim of helping institutions to identify and repair gaps more quickly.
European sovereignty AI discussion on warming
The European discussion on “Sovereign AI” rose rapidly last week after the United States suddenly cut off access to the Anthropic front model. For European enterprises and policymakers, the risk of disruption of supply of external models is no longer theoretical.
Cohere CEO Aidan Gomez stated that sovereignty AI should be based first and foremost on locally controlled infrastructure, including chips, electricity, data centres and privatization deployments. If it was difficult for individual countries to do so on their own, the reliance on the infrastructure of the United States and China should be reduced through strategic alliances.
However, Amazonian views are more pragmatic. DeSantis argues that few countries have a fully independent AI infrastructure and that it is more realistic to leave sensitive data locally and ensure that governments or businesses have clear control over the use and governance of AI, rather than re-establishing complete hardware and supply chain systems in each country.
The company started asking for ROI.
Another change on the floor was the decline in business executives' interest in “dazzling demonstrations” and the shift to more training, process automation and cost control. Philip Rambach, Chief Executive Officer, Schneider Electric, stated that the company had requested 42,000 employees to receive AI training and only supported projects that had a clear business landscape and could scale up applications.
He stated that the company wished to use AI as an operational tool, rather than just at the level of innovation. This statement also reflects the fact that large enterprises are tightening AI pilot standards, with greater emphasis on operational effectiveness.
A similar change was observed in OpenAI. According to its senior management, business clients still want to use the strongest model, but the more frequently asked questions have become: Whether or not these models actually provide returns and whether the deployment of a large number of AI parties has produced sufficient value.
One of OpenAI ' s strategies to respond to such needs is to continue to reduce the cost of using models, to improve efficiency and to enable enterprises to do more with lower inputs. The article shows that, as AI moves into a larger phase of deployment, costs, control and security are gradually replacing the pure technological heat to become a core standard for enterprise procurement and policy discussions.