AI Watch, Tuesday 29 September 2026Your agents also serve those attacking you

The French unicorn is setting up offices in Munich, close to Germany's industrial base, to bring its teams nearer to large manufacturing and automotive accounts. The geography states the strategy: after making its mark on open models and sovereignty, Mistral is looking for deployments in European industry, where the budgets and the heavyweight use cases are. The German base complements the Paris roots rather than replacing them.

The encyclopedia's community has moved the use of text-generation models from "strongly discouraged" to a general ban, with only two exceptions. The decision aims to protect the reliability of articles against the fabrications and plausible errors these tools produce, in an encyclopedia where every claim must point to a verifiable source. The move stands apart from the drive to embed AI everywhere else, and draws a clear line for one of the most visited sites in the world.

Fraudsters extracted €95 million from Fideuram, the private-banking arm of Intesa Sanpaolo, according to Numerama. The scheme combined two levers: impersonating the boss on WhatsApp, then a call in which a lawyer's voice, cloned by AI, approved the transfers. More than a third of the sum is still missing to date. The case shows that CEO fraud, a known technique, becomes formidable when the voice giving the order is one you think you recognise.

The American company unveiled on Monday 28 September the new version of its mid-range model, presented as a work partner "significantly cheaper and faster". Sonnet 5.5 runs more than 30% faster than Sonnet 5 and costs up to 30% less per task, at an unchanged price per token (the unit of text that is billed). On Terminal-Bench 4.0, a coding test, it jumps from 10.3% to 70.6% and beats Opus 5.5, the company's high-end model. The market logic holds: the best performance-to-price ratio no longer comes from the heaviest model, but from a mid-range model sharpened for everyday tasks.

The company has shelved the GPT-6.1 Astra version of its model, which scored worse on safety evaluations than the version already in service, Bloomberg reports. The new build did improve on the model's "laziness", its tendency not to finish a requested task. An executive told the Wall Street Journal that the model followed instructions poorly. The decision comes on the eve of OpenAI's developer conference, an event built to show off new releases, and extends a run of cautious moves after several safety incidents this summer. Holding back a more capable model because it fails safety tests remains rare in a sector chasing power.

Jensen Huang presented on Monday 28 September the Open Agent Safety Platform, a set of software and hardware that adds independent safety layers around AI agents. The tool, presented in response to a wave of incidents where agents breached the walls meant to contain them, promises to isolate within milliseconds an agent trying to step outside its perimeter, including during execution. Nvidia is betting on open code to make it a shared building block for the sector. The chip supplier that fed the wave of agents now also offers the harness to hold them.

The firm Bain & Company puts at $6 trillion the annual revenue the AI industry will have to reach by 2031 to justify the capital sunk into building data centres. Productivity gains and consumer uses alone will not bring in enough, warns the report relayed by La Tribune: new markets will have to emerge to close the gap. The figure gives a measure of the bet: $6 trillion a year, more than the annual gross domestic product of Japan, the world's third-largest economy, to be found every year just to cover the infrastructure bill.

The owner of TikTok accounts for about a fifth of the data-centre capacity already delivered in China, making it the country's biggest tenant and the main driver of its rise in AI, according to new estimates from the firm SemiAnalysis. The figure rests on tracking more than 1,000 facilities operated by more than 60 companies. ByteDance rents almost all of its fleet rather than owning it, and uses that power to feed its AI products, including the Doubao assistant. The concentration illustrates a Chinese model in which a few digital giants alone pull demand for compute.

Samsung Electronics and five of its subsidiaries are pouring a billion dollars in total into an American company run by a former Amazon Web Services executive and backed by the KKR fund, Bloomberg and Nikkei Asia report. The deal reinforces the South Korean group's push into the fast-growing AI-infrastructure market. A maker of memory and chips, Samsung is also positioning itself as an investor in the centres that will consume those components, securing an outlet for its production. South Korea intends to count in a race where compute is becoming the strategic asset.