k-flash post

Patrizia Ascari

ABOUT THE AUTHOR

Patrizia Ascari is holder of a bar admission, as well as law degrees from the University of Geneva (Switzerland) and Università degli Studi di Firenze (Italy) and a postgraduate in corporate tax law from Bocconi University in Milan (Italy). In recent years she has contributed to the implementation of FATCA and the Automatic Exchange of Information (AEOI) within the banking sector. Patrizia has also had the privilege of serving on several occasions as a speaker at SwissBanking seminars and part of the team of trainers at VisionCompliance. Patrizia is also a recognized yoga teacher with Black Lotus Yoga. After ten years with a renowned Geneva-based private bank, she voluntarily stepped down from her position to devote herself fully to yoga, which she teaches regularly in Geneva. As part of this journey, Patrizia is currently studying Hinduism at the Oxford Centre for Hindu Studies. In parallel, and in a path that is perhaps not entirely disconnected from yoga, she is attending the Certificate of Advanced Studies (CAS) in Asylum Law at the University of Fribourg.
EPFL explores workaround to mitigate risks of AI agents bypassing instructions

by Patrizia Ascari | Aug 27, 2026 | Information Technology/AI

k-flash audio commentEPFL explores workaround to mitigate risks of AI agents bypassing instructions
k-flash audio commentEPFL explores workaround to mitigate risks of AI agents bypassing instructions

The case

Autonomous AI agents are becoming the longer the more capable of bypassing their safety safeguards and carry out unauthorized actions on their own.

Source: Le Temps, EPFL – AI Sécurité, 21.08.2026

The commentary

EPFL researchers have shown that these systems can be manipulated by breaking straightforward harmful objectives down into a series of smaller seemingly harmless steps.

In a simulation environment, EPFL tested this approach on models such as GPT, Gemini, Claude and found out that this approach can be surprisingly effective. To make sure such risk is reduced, the researchers emphasize the importance of building robust safety measures much earlier, while developing multi-agents systems. Measures must be built into AI systems from the outset, rather than being added only after problems arise.

This publication has been prepared solely for information purposes and is does not constitute a recommendation, a solicitation, or an offer. The information on which this publication is based has been obtained from sources that we believe to be reliable and in good faith, but we have not independently verified such information and no representation or warranty, express or implied, is made as to its accuracy. All expressions of opinion are made as of the date of publication and may be subject to change without notice. k-flash and all related affiliates accepts no liability or responsibility whatsoever for any consequential loss of any kind arising out of the use of this publication or any part of its contents. The use of this publication should not be regarded as a substitute for the exercise by the recipient of his or her own judgment. This publication is not directed to any person in any jurisdictions that prohibit such publication.
Click to rate this post!
[Total: 1 Average: 5]

RECENT POSTS

Fischer Convicted, Xhaka Under Investigation

Sep 22, 2026 |

The Case: In 2023 Patrick Fischer, former head coach of the Swiss national ice hockey team, was convicted of document forgery in connection with a falsified Covid vaccination certificate and was given a monetary penalty of CHF 38,910. Now another case involving a famous sportsman has evolved. The Commentary: The…

Something beyond happiness

Sep 20, 2026 |

The Case: Moments that many regard as the height of happiness are, for some philosophers, merely fleeting episodes of pleasure. Once they fade, the restless mind begins reaching again for new goals, experiences and distractions. The Commentary: For Schopenhauer, many of the experiences commonly associated with happiness – travelling, golden…

Search

TRANSLATE

Newsletter

Register Now

Related Posts

The AI Elite vs. Magnifica Humanitas

Jul 13, 2026 |

The Case: This thought-provoking book brings together the people shaping the future of artificial intelligence. The pioneering scientists who laid its foundations, the researchers and engineers driving today’s breakthroughs, the leaders behind OpenAI, Google DeepMind, Microsoft, Nvidia, Meta, Anthropic, SpaceXAI, Amazon and Oracle, the investors pouring billions into the AI…

Apology for Duplicate K-Flash Publications

Jul 6, 2026 |

The Case: What will we remember about 2024? The Commentary: Through a selection of photos from the Swiss media, the exhibition “Swiss Press Photo 25” reveals the prizewinning images from the competition of the same name that celebrates the best of Swiss photojournalism every year.

Strategy for National AI infrastructure

May 18, 2026 |

The Case: The Swiss Science Council (SCC) sees significant strategic opportunities for Switzerland regarding data research and compute-intensive research. The Commentary: In their most recent report, the SSC highlights the potential for the further development of existing academic computing infrastructures in a targeted and sustainable way so growing demands can…

Categories