On May 12, in a Brazilian labor court, a legal appeal contained a white text line on a white background commanding the software to review the case superficially. The software itself detected the injection. We covered the case in detail here.
The question for practitioners in Italy is how long before something similar becomes possible here.
Brazil and Italy: Current Status
| Brazil | Italy | |
|---|---|---|
| National System | Operational across all regional labor courts | Expected by August 2026, not yet operational |
| Pilots & Trials | Completed, in full production | Catania with CINI, Cremona Preliminary Judge office |
| Digitization of Filings | Completed | NRRP Data Lake project in progress |
| First Documented Attack | May 2026, detected and sanctioned | No documented equivalent case to date |
Italian Modernization by the Numbers
€133 million total NRRP investment allocated to justice digitization, including the Data Lake.
7.75 million case files to be digitized across Supreme Court, appellate courts, tribunals, and prosecutors.
6 integrated knowledge systems planned to turn court records into structured data ready for AI analysis.
Overseeing these projects since 2024 is a permanent Ministry observatory tasked with defining operational guidelines. What remains missing is the core piece: a unified national platform within the justice domain. The High Council of the Judiciary (CSM), in its resolutions, anticipated its availability before August 2026, yet recent analyses confirm it is not yet operational. In the interim, official guidance recommends using only authorized internal tools, as public consumer online services lack required privacy and data governance safeguards.
The Regulatory Timeline Has Shifted
Annex III, point 8 of the EU AI Act classifies AI systems intended for use by judicial authorities to assist in researching and interpreting facts and the law, or applying the law to specific cases, as high-risk. This is not an automatic label for every court software: tools performing narrow procedural or preparatory tasks, or refining work already completed by a human, are excluded.
The Italian national roadmap was initially calibrated around August 2, 2026. However, the Digital Omnibus package pushed the Annex III deadline to December 2, 2027.
The Paradox
Italy had not finalized its transition by the original deadline, and that deadline has now moved back by sixteen months. This creates unexpected breathing room, and the outcome will depend entirely on how this window is utilized.
The Ingestion Channel Is Already Open
In Italy, court filings have been submitted electronically for years. Documents already enter the infrastructure in digital form; only the automated reader on the other end is missing. When it goes live, incoming files immediately become an entry vector into a text-processing language model.
Judicial offices face a unique constraint. While a private company can reject an unusual resume or mandate a specific format, a court is legally required to accept filings submitted by parties. An attacker does not need to breach a perimeter; exercising a procedural right is sufficient.
Decisions to Make Today
Architects building systems that read third-party documents today can enforce a foundational rule that is very difficult to retrofit later: treating external documents strictly as data to inspect, never as instructions to execute. In practice, this comes down to three technical decisions.
Three Upfront Architecture Decisions
1. Keep only human-visible text. During text extraction, preserve only what a human reader sees; route hidden or suspicious layers into an audit report rather than silently discarding them.
2. Explicit provenance tagging. Explicitly instruct the model that extracted text represents a party's assertion and contains zero valid execution commands.
3. Log and sanction attempts. A system that silently filters attacks protects individual cases without establishing deterrence. The Brazilian ruling imposed real financial costs on the attempt, which is what shifts long-term behavior.
This applies equally to courts and to any business feeding external files into language models: invoices, RFPs, resumes, medical records, or contracts under review. This is the domain we address with TCT, where agent inputs are treated as untrusted by default, and with Guardiano EU for regulatory compliance.
Sources & References
- Official statement by the Regional Labor Court of the 4th Region (TRT4) regarding the Galileu system, case details, and incident date. Sanction details reported across Italian legal publications including Il Sole 24 Ore (NT+ Diritto), Altalex, and Agenda Digitale.
- Italian Ministry of Justice, NRRP digitization investment (M1C1-I1.6.2), Data Lake architecture, and case file digitization targets.
- Lexintel project by the Court of Catania with CINI consortium and local Bar Association; "Augmented Justice" pilot at the Cremona Preliminary Investigation Judge office.
- Regulation (EU) 2024/1689, Annex III point 8, and Regulation (EU) 2026/1744 deferring application to December 2, 2027.
- On national platforms and High Council of the Judiciary resolutions, analysis by Giovanni Pascuzzi: Intelligenza artificiale e sistema giudiziario: quella piattaforma che ancora non c'è (August 11, 2026) and Le delibere del CSM sull'intelligenza artificiale nell'amministrazione della giustizia (July 27, 2026).