The proof of the machine: verifying AI that acts | Radar Weekly N12

Radar Weekly N12: a golden computing cube and luminous traces crossing verification panels, with the official Gruppo Radar signature.

The proof of the machine is not only a convincing result: it is a path that can be reconstructed and checked. Radar Weekly N12, dated 7 September 2026, compares agentic actions, formal verification and infrastructure to understand when trust in AI becomes genuinely verifiable.

The test of the machine: verifying AI that acts

Weekly Number 12 | Monday, 7 September 2026 | Approximately 6 minutes to read

1. After the result, what remains visible?

A system can produce a correct answer without making clear how it arrived there. If it then uses tools, modifies data or accesses other services, the final result tells us even less: it does not automatically show which permissions were used and which steps took place.

The issue follows this difference through very different cases. A side channel used during a cyber evaluation, a proof translated into a formal language, a platform whose ownership changes and a European threshold for responsibility pose the same question: who can scrutinise the conditions of the action?

This is not about attributing what we might colloquially call “human intentions” to models. Goals, credentials and environments are defined by those designing and deploying them. Precisely for this reason, the record of work is not an administrative detail: it distinguishes permitted activity from an unauthorised step.

2. A side door can lie in ordinary tools

In the reconstruction of the Hugging Face incident, OpenAI describes research models used in internal cyber evaluations in July 2026, with reduced safeguards. This is not an account of ordinary public use of ChatGPT.

A service intended for software-package management is used as an unintended communication channel. The instructive point is the environment: an ordinary component can connect agents, tools and networks in ways the initial design had not considered.

Security must therefore also examine the possibilities offered by intermediaries. Asking a model to respect a rule does not replace checking the access actually available. Declaring a boundary is useful only if the tools do not allow it to be crossed by another route.

Even a correct record of the anomaly is not enough. Someone must interpret it, assess its extent and have the mandate to suspend the sequence. Without this possibility, control risks arriving after the effect has already occurred.

🛡️ The sandbox is a boundary to verify

A sandbox is an environment limiting what a program can access. Logs are records of activity. The former should contain the action; the latter help reconstruct it. Neither replaces those who must read alerts, intervene and check that a correction works.

3. A formal proof changes the kind of trust

On 4 September, Anthropic presents the formalisation of Fermat’s Last Theorem in Lean. The point is not to claim that AI discovered the solution to a problem already proved in 1995, but to make a proof checkable by a software verifier.

Lean is a proof assistant: it checks a logical chain against explicit rules. A formalisation must express the steps in a language the verifier can examine, without relying on the persuasiveness of the account.

The difference from a plausible answer is concrete: there is an object to submit for checking; judgement about the assumptions, correspondence with the problem and meaning of the result remains human.

The contrast between the two cases helps prevent trust in AI from being discussed as though it were a single quality. An action crossing an unintended channel and a path built for verification require different assessments.

🔎 Checking logic does not mean judging everything

A formal verifier can check whether a conclusion follows the declared rules. It does not decide on its own whether the chosen problem matters, whether the assumptions describe a real situation well or whether the result is useful. A checkable proof reduces one area of uncertainty; it does not replace every form of judgement.

4. A platform’s openness must be checked too

Clarity also concerns terms of access. In the announcement of the agreement to acquire Hugging Face, NVIDIA states that the platform will remain open and usable with multiple clouds and accelerators. At the issue’s date, this is an announced agreement, not a transaction already completed.

For those using models, data and evaluation tools, the question does not end with a promise of openness. The ability to choose a computing provider, move work and maintain access suited to one’s needs matters.

Interoperability allows different systems to communicate and work together. Equally important is being able to choose among several providers, without the whole process depending on a single company. To understand how real this openness is, we must examine how services work and the contractual terms: that is where we see which tools can be connected, used alongside one another or replaced over time.

Alongside this question, the dossier also considers server orders, backlog and data-centre projects. A backlog is a queue of orders still to be fulfilled: it signals work booked, not capacity delivered or guaranteed future revenue. Here too, precision prevents promise from being confused with availability.

5. Scale takes verification outside the laboratory

On 31 August, the European Commission designates ChatGPT as a very large online search engine under the Digital Services Act. The stated threshold is at least 45 million average monthly users in the European Union.

The designation does not establish harm, but links the service’s size to additional risk-assessment and mitigation obligations. The issue gives January 2027 as the expected deadline for compliance.

Responsibility thus meets users, procedures and authorities. Declaring a system safe is not enough: the processes adopted and measures applied must be open to scrutiny. The same need runs through technical research and the distribution of a service at large scale.

6. The test of the machine remains a human responsibility

Who assigns the goal? Who grants the tools? Who can check the result, challenge it and stop the process? Separating these questions prevents the whole chain from being compressed into the phrase “AI did it”.

The issue’s comparison does not lead us to seek trust, nor to regard every automation as unreliable. It leads us to seek inspectable paths, effective limits and evidence others can check.

Responsibility remains with those developing, using and governing these systems: people, companies, laboratories and institutions. Exercising it requires knowing what happened, establishing who can access different functions and having the tools to intervene. Careful design makes these steps possible, preserving useful records and effective controls. The test of the machine begins when we can compare what it promises with what it does, examine its results and assess its consequences.

Read the full issue

This reading presents some stages of Radar Weekly N12. The full issue contains the complete journey, data, in-depth analysis and sources of the dossier of 7 September 2026.

© 2026 Barbara Gattobigio · Gruppo Radar
Comprendere oggi. Costruire domani · by effettoantropico.life
All rights reserved on original content. Free access does not authorise reproduction or republication without consent, except as permitted by law. Third-party materials remain the property of their respective owners. To share this issue, use the link to the article.

Continue reading

Understanding Today. Building Tomorrow: the project and the Radar collection.

Cover in focus

proof - English cover of Radar Weekly N12, The proof of the machine, dated 7 September 2026: a luminous cube and traces crossing transparent verification panels, with the official Radar signature.

Read this article in Italian