Spain exempts public bodies from AI Act fines, and AI leaks your phone number
Spain's government quietly exempts public administrations from the EU AI Act's biggest fines, while Spanish shopkeepers describe chatbots handing out their phone numbers by mistake. Plus Apple joins the AI agent race and Mistral expands into Asia.
Every line links to a primary source in the full briefing.
01Spain exempts its own public administrations from the EU AI Act's heaviest fines
Spain's government has moved to shield public administrations from the largest penalties under the EU AI Act for noncompliance. The carve-out means enforcement will fall more heavily on private companies than on the government bodies that also deploy AI systems.
Business owners now have a concrete reason to expect enforcement to focus on them rather than on public bodies using similar systems, which sharpens the case for getting compliance paperwork in order.
Source: Cinco Días
02Chatbots are handing out real people's phone numbers to random callers
A Spanish report describes people receiving a stream of calls meant for completely unrelated businesses, from bike shops to clinics to pharmacies, after Gemini and ChatGPT mistakenly offered up their personal mobile number as a contact point. The affected individuals have no connection to any of the businesses customers are trying to reach.
A shop that lists its contact details online has no guarantee an AI assistant will quote them correctly, so wrong numbers circulating in chatbot answers can mean lost customers calling someone else entirely, or your own number ending up linked to a stranger's business.
Source: El País
03Apple reportedly building its own AI agent for the Mac
A job listing spotted by Bloomberg suggests Apple is developing an AI agent that can browse the web, work across apps, run commands and write code on a user's behalf, similar to agents already shipped by Meta, Anthropic and OpenAI. Apple has not confirmed details or a timeline.
If every major platform ships its own agent that can act inside your software and accounts, you will eventually need a policy on which agents are allowed to touch company systems, well before any of them land on your employees' machines.
Source: TechCrunch
04Mistral opens an Asia-Pacific office in Singapore, aiming for 200 staff by 2027
French AI company Mistral is setting up a regional headquarters in Singapore and plans to grow that office to 200 employees by 2027. The move extends Mistral's presence beyond Europe as it competes for enterprise customers against US and Chinese labs.
A stronger Asia-Pacific push from a European lab gives businesses outside the US a credible non-American alternative to weigh for data residency or contractual reasons when choosing a model provider.
Source: KuCoin
05SAS flags three recurring problems enterprise AI buyers will face into 2027
Analytics company SAS has outlined three challenges it expects to keep tripping up companies deploying AI going into 2027, based on patterns it has seen among its enterprise customers. The piece does not detail novel technology, but reflects where real deployments are getting stuck.
Hearing the same adoption problems named a year in advance is a cue to budget time now for governance and data readiness instead of discovering the gaps mid-rollout.
Source: El Ecosistema Startup
06OpenAI publishes a reference setup for supervised AI security investigations
OpenAI released a reference architecture called Daybreak Red for running AI-assisted security investigations under human supervision, including safety controls and monitoring guidance. It is aimed at security teams wanting to use AI models on sensitive internal investigations without losing oversight.
Any company letting AI tools near security logs or incident response should treat human sign-off as the default, not an afterthought, and this gives a concrete pattern to copy rather than build from scratch.
Source: OpenAI
07A new provider lets the open source Kimi coding agent use AWS credentials directly
Kimi Code CLI, an open source coding agent, now has a native Amazon Bedrock connection that uses a company's existing AWS sign-in (including single sign-on) instead of requiring a separate API key to manage. Previously, connecting the agent to Bedrock meant creating and protecting an extra credential.
Fewer stray API keys floating around means one less thing for an IT team to lose track of when adopting AI coding tools, which matters more than the specific tool chosen.
Source: AWS Machine Learning
Check who is legally on the hook for AI mistakes in your own country
With Spain's government carving public bodies out of the EU AI Act's heaviest fines, enforcement pressure is shifting toward private businesses. If you operate in Spain or elsewhere in the EU, spend an hour this week confirming who in your business is responsible for AI Act compliance paperwork, since regulators now have one less target to go after first.
Sources
7 links- 01El Gobierno libra a las administraciones públicas de las multas millonarias por incumplir la ley europea de IACinco Días
- 02Llamadas para comprar una bici, mangos o un tratamiento para la vagina: qué pasa si la IA distribuye tu número de teléfonoEl País
- 03SAS anticipa 3 desafíos de la IA empresarial en 2027El Ecosistema Startup
- 04Use Kimi Code CLI with models on Amazon BedrockAWS Machine Learning
- 05Deploying Daybreak RedOpenAI
- 06Mistral AI Opens APAC HQ in Singapore, Aims to Hire 200 by 2027KuCoin
- 07Apple may be developing its own AI agent that runs your MacTechCrunch