TL;DR
Since 30 September 2026, Italy has a criminal offence aimed specifically at AI security failures.
A new Article 437-bis of the Penal Code punishes two things: failing to adopt the technical security measures prescribed for high-risk AI systems, and unlawfully altering such a system. Both require that the conduct give rise to an actual danger — to life, to individual or public safety, or to the security of the State. Penalties run to eight years for the omission offence and ten for the alteration offence at the State-security tier.
A companion article brings this inside Italy’s corporate liability regime, so an organisation can be sanctioned alongside any individual.
What it is not: a general duty to secure AI systems enforced by criminal law. The danger element is a real threshold, and a failure that endangers nobody does not engage the provision.
Where It Sits
The instrument is secondary legislation: a legislative decree made by the Government under a delegation granted by Italy’s AI law of 23 September 2025. It is not an EU instrument and it is not an implementation of the AI Act’s own penalty regime, though it uses the AI Act’s concept of a high-risk system as its subject matter.
Three dates, which should not be conflated:
| Signed | 9 September 2026 |
| Published | 15 September 2026 — Gazzetta Ufficiale, Serie Generale no. 214 |
| In force | 30 September 2026 |
The Offences
Article 437-bis, first paragraph — omitting security measures
The conduct: omitting the technical security measures prescribed for the design, training, production and placing on the market of high-risk AI systems — measures suitable to prevent malfunctions or alterations of the systems’ operation.
The threshold: the omission is punishable where a danger to life or to individual or public safety results, carrying one to five years. Where the danger is to the security of the State, two to eight years.
Read the elements in order, because the order matters. There must be a prescribed measure; it must be one directed at preventing malfunction or alteration; it must have been omitted; and a qualifying danger must have resulted. A deployer who can show the measures prescribed for its system were adopted has answered the charge at the first element.
Article 437-bis, second paragraph — unlawful alteration
The conduct: outside the cases in the first paragraph, altering a high-risk AI system. The paragraph opens with the standard Italian subordination clause — it applies save where the act constitutes a more serious offence.
The same danger element applies: two to six years where danger to life or individual or public safety results, and three to ten years where the danger is to the security of the State.
Note the placement in the code. Article 437 concerns the omission of precautions against workplace accidents or disasters — so the legislator has sited AI security alongside industrial safety offences rather than alongside computer crime. That is a signal about how the provision is meant to be read.
Corporate Liability: Article 25-vicies of D.Lgs. 231/2001
Italy’s Legislative Decree 231/2001 makes entities liable for offences committed in their interest or to their advantage, with a defence available where an adequate organisational and management model was in place and effectively implemented.
The decree inserts a new Article 25-vicies, headed offences committed using artificial intelligence systems. For the Article 437-bis offence an entity faces a pecuniary sanction of 600 to 1,000 quote, alongside the disqualification sanctions available under the regime.
This is the provision that will change behaviour. The 231 regime is already how Italian corporate compliance is organised, and bringing AI security inside it means AI controls now belong in the organisational model that boards already maintain, review and are advised on — rather than in a separate governance document nobody litigates over.
What It Does Not Do
- It does not create a general prohibition on solely automated decision-making. Narrower provisions on that theme appear in the part of the decree dealing with police and law-enforcement use of AI, including on the labelling, filtering and categorisation of biometric datasets. Nothing in the criminal or civil-liability parts creates a general rule.
- It does not criminalise a security failure as such. Without a resulting danger to life, safety or State security, Article 437-bis is not engaged.
- It does not define the technical security measures itself. The offence refers to measures prescribed elsewhere. What is prescribed, and by what instrument, is the question an Italian deployer has to answer first — and it is the one that will generate the early litigation.
- It is not the AI Act’s penalty regime, which is separate and unaffected.
What an Organisation Operating in Italy Should Do
- Establish whether any of your systems are high-risk within the meaning used by the decree. Everything follows from that, and nothing below applies if the answer is no.
- Identify the prescribed technical security measures that attach to those systems, and the instrument that prescribes them. Record the analysis.
- Evidence adoption, with dates. The first element of the offence is omission; a dated record of adoption is the direct answer to it.
- Bring AI into the 231 organisational model. Treat this as you would any other predicate offence: risk mapping, controls, the supervisory body’s remit, training, and a documented review cycle.
- Map the danger element onto your own systems. Which of your AI systems could, if they malfunctioned or were altered, endanger life, public safety or State security? That list is short for most organisations and it is where the criminal exposure actually sits.
- Control change. The second offence is alteration. Who can change a high-risk system’s behaviour, under what authorisation, with what record? That is the same question as human-in-the-loop design asks, now with a criminal consequence attached.
The underlying assessment work is in pre-deployment risk assessment; the EU-level obligations this sits beside are in EU AI Act compliance.
Status and Limits of This Note
The provisions above were read in the Gazzetta Ufficiale text and checked against the consolidated Penal Code and 231/2001 texts on Normattiva, which show both new articles in force from 30 September 2026.
This is a summary of the instrument’s structure and thresholds, not legal advice, and the Italian text governs. No prosecution under Article 437-bis is known to have been brought, so there is no judicial interpretation of the danger element, of what counts as a “prescribed” measure, or of how the alteration offence interacts with Italy’s existing computer-crime provisions. Those are the three questions on which early practice will turn.