This article is devoid of information. Go to the source: https://www.etsi.org/deliver/etsi_tr/104100_104199/104171/01.01.01_60/tr_104171v010101p.pdf

This is basically the snippet of the source that they’re writing the article about:

AI attacks against QRNGs are a form of side-channel attack or predictive analysis. They do not try to break the
quantum laws of physics but rather exploit the flow of information between the quantum source and the final output.

Exploiting Classical Noise: Every QRNG system has classical hardware components, like sensors, power supplies, and
signal processors. These components introduce small amounts of classical noise, which may be predictable in some
cases. If these untrusted noise sources have not been included in the QRNG’s physical model (see Definition 3), an
attacker with a sophisticated AI model may be able to analyse the output of the QRNG and, over time, learn to identify
the predictable patterns caused by it. Then, once the distribution of the classical noise is known, the attacker can now
slightly improve its ability to predict the QRNG’s outcome, thereby compromising the security of the generated random
numbers. Even for an advanced AI, this can be a fairly difficult and time-consuming procedure.

Side-Channel Attacks: These involve an attacker monitoring side channels of the device, such as power consumption,
electromagnetic radiation, etc. AI can analyse extremely large amounts of data fairly quickly and potentially find
correlations between these side-channel signals and the output of the QRNG. This allows an attacker to gain
information about the output without directly attacking the quantum source. For example, a QRNG might show a subtle,
non-random power spike when it generates a specific type of number, a pattern that an AI could learn and exploit. This
is not unique to QRNG attacks and may be applied in a similar manner to other types of random number generators as
well.

To counter these threats, developers of QRNGs are focusing on a few key strategies:
• Robustness against Classical Noise: By improving the hardware design and shielding of QRNG devices,
manufacturers can minimize the impact of classical noise.
• Advanced Randomness Extraction: Researchers are developing more sophisticated and verifiable
randomness extraction algorithms.
• Continuous Monitoring: As previously discussed in clause 5.1.4.3, continuous monitoring systems that
constantly check the QRNG’s output for even the slightest biases or correlations may allow for real-time
adjustments or alerts.

It’s nothing new. Just reiterating that implementation details matter in crypto systems.