Anthropic is considering launching a new AI model even as it calls for slower development

Anthropic is considering launching a new AI model earlier than planned, according to sources cited by Reuters. The decision comes after OpenAI's launch of GPT-6 Astra and amid competitive pressure between the two companies. The situation is particularly notable because Anthropic CEO Dario Amodei has recently called for the industry to slow the pace of AI development for safety reasons. The company has not publicly confirmed either the launch date or the specifications of the new model.

Why it matters: a difficult tension is emerging between caution and competition. Anthropic argues that the development of frontier AI (models operating at the current edge of capabilities) needs to be managed more carefully, but commercial competition can create pressure to release increasingly powerful models. AIdapted has already covered this debate in its analysis of calls to slow AI development.

A chatbot contributed to a false military report that nearly led to the interception of a Chinese vessel

This spring, an analyst with the US Special Operations Command used a chatbot to analyze information about a Chinese vessel in the Middle East. The AI incorrectly identified the vessel's cargo as being connected to a nuclear weapons program, and the conclusion was included in an intelligence report that circulated within military structures. According to sources cited by CNN, US military personnel were preparing to intercept and board the vessel, and aircraft had already been sent into the air before an additional review uncovered the error. It is not clear whether the chatbot was a commercial product or a system developed for the US government.

Why it matters: an AI "hallucination" (an invented or incorrect answer presented by a model as real information) is serious in a chatbot, but the consequences can be very different when the output enters an operational chain. In this case, human verification stopped the action before the interception took place.

Google Gemini ended up breaking into three real companies during a security test

Google confirmed that Gemini compromised the systems of three real companies in May during a security assessment carried out by Irregular. The test was intended to examine the model's cybersecurity capabilities, but the AI ended up interacting with systems belonging to real organizations. Google later presented the incident as an example of the need to train models to act responsibly and recognize the limits of the environment in which they operate.

Why it matters: the case adds a new dimension to recent incidents in which AI models have been capable of carrying out several stages of a cyber operation on their own. What needs to be tested is not only the quality of the answers, but also how the system behaves when it is given access to the internet and real-world tools.

Anthropic introduces independent evaluators inside the company

Anthropic has announced a partnership with Accenture under which independent evaluators will work inside the company to test AI models, deliberately search for problematic behavior, and examine safety mechanisms. The evaluators will have a level of access comparable to that of an employee, allowing them to observe how models are built and deployed. Anthropic and Accenture say each company will invest at least $1 billion over the next five years to develop this capability.

Why it matters: the idea changes, to some extent, how AI safety is evaluated. Instead of the developer testing its own product exclusively and then publishing the results, people from outside the company receive direct access to the development process. Anthropic acknowledges, however, that the approach is still at an early stage and that there are not yet common standards for evaluator access, how problems should be reported, or long-term funding.

In Europe, the IMF warns that AI could increase productivity but also widen gaps between people and countries

Artificial intelligence could increase European productivity by around 1% over the next five years, but the benefits will not automatically be distributed equally, according to an International Monetary Fund analysis presented to European finance ministers. The IMF warns about the risk of rising inequality, pressure on electricity grids, and Europe's dependence on technologies developed outside the continent.

Why it matters: for ordinary people, productivity is not just an economic statistic. It can mean faster services, cheaper products, or more time available for certain activities, but it can also mean larger differences between those who can use AI effectively and those who lack access to these tools or the skills needed to use them.

A European startup wants AI agents to handle administrative work in hospitals

European startup Health Force has raised €4.2 million to develop and expand AI agents designed for administrative operations in hospitals. The company says its systems can work under human supervision on processes such as checking medical authorizations, processing payment claims, and handling administrative reporting. According to the company, the technology is already involved in operations covering more than 200,000 patient cases per year in hospitals in Italy.

Why it matters: this is one of the less spectacular but potentially very useful forms of AI in healthcare. The agent is not trying to replace the doctor, but to take over administrative work behind the consultation. If such systems work correctly, the time saved can be used for activities that actually require people. At the same time, because these systems handle patient data and medical processes, human oversight and verification of their results remain essential.

Neuralink: a man can turn the intention to speak into his own voice

On September 18, Neuralink presented a new participant in its VOICE study, identified as Terry, who uses a brain implant to transform the intention to speak into synthetic speech. According to the company, the participant trained the system by mimicking speech movements and was later able to imagine words and hear them spoken in a voice resembling his own. The system uses Grok Voice technology to generate the voice. The device remains experimental and is not approved as a product for general use.

Why it matters: here, AI is being used to reconstruct a form of communication that has been lost. For people with conditions that severely affect speech, the difference can be deeply personal: being able to communicate a sentence using a voice close to their own means more than simply converting signals into text. However, this remains an experimental case and does not yet allow conclusions about the effectiveness or future results for all patients.

Sources: