Following this week's incidents with test agents at Anthropic, Microsoft chief executive Satya Nadella is weighing in on the debate about controlling AI. His advice is concrete enough to hold up against your own AI use. We also look at our own data: in the arena leaderboard, more changed last month than in the months before.
Nadella: assume a model can go off the rails
On Saturday 10 October, Nadella wrote on X that it is time to revisit the "trust architecture" of AI, TechCrunch reports. His starting point: "We must assume a model is compromised and contain it from the start." He compares this to an emergency brake.
According to TechCrunch and the Taipei Times, his advice boils down to these points:
- An authorised person must always be able to pause or stop a model in the middle of a task.
- Keep the model separate from the software that directs its work, and place controls outside the model.
- Record every significant action of a model in tamper-proof, readable logs.
- Do not rely on a single model for critical decisions.
- Have AI systems independently audited, and make major errors or breaches public.
Nadella is responding to a series of incidents, including the report that Anthropic was unable to reliably control its test agents. It is an opinion piece on X, not a new Microsoft product or policy.
What does this mean for you?
You do not need to be an AI lab to apply this. If you let an agent send emails, fill in forms or look up client data, ask yourself three questions. Can you stop it while it is running? Can you see afterwards exactly what it did? And does an outgoing action first go past a human? A log of agent actions also helps if you have to explain under the GDPR what happened to personal data. If you work at a bank, an insurer or in government, then "don't rely on a single model" for critical decisions is a sound starting point.
The arena in a month: Gemini jumps, Mistral and DeepSeek climb
We track the highest LMArena score per model family every day. Compare the standings for 11 October with those for 11 September and three things stand out:
- Gemini went from 1494 to 1525 points (+31). That jump came all at once on 2 October, with Gemini 4 Argon (high). It now sits in place 1, 18 points ahead of Claude Opus 5.5 (high) on 1507. Note: Argon has only 4892 votes and a margin of ±9 points.
- DeepSeek rose from 1459 to 1475 (+16). The top model is deepseek-v4.1-flash-max, with an MIT licence and $0.30/$1.20 per million tokens (input/output).
- Mistral rose from 1414 to 1429 (+15), with Mistral Large 4 at $0.68/$2.09 per million tokens. It sits in place 115 of 414, still well outside the top.
Claude stayed level on 1507, ChatGPT went from 1483 to 1485, and Kimi dropped one point to 1488.
How high a model ranks says nothing about what happens to your data. DeepSeek scores well and is cheap, but it is offered from China. Mistral is a European provider. On /veiligheid/ you can see how each model handles your data. Compare price and quality at /modellen/.
In brief
- Nadella wants an emergency brake, logs, multiple models and external audits for AI agents. It is advice, not policy.
- Gemini 4 Argon has been at the top of the arena since 2 October, but with relatively few votes.
- DeepSeek and Mistral climbed the fastest this month after Gemini. Check the GDPR profile of both before you deploy them.