| Homepage | https://nicolaszahn.ch |
The first cyberattack in history using prompt injection. Attackers used Meta’s chatbot as a tool to take over Instagram accounts belonging to well-known people, brands, and institutions.
Basic mistake: using LLM as a security boundary.
https://www.theverge.com/tech/941179/meta-instagram-ai-support-chatbot-exploit-hacked
more and more projects are adding an AGENTS.md that explicitly tells agents to stop, drop, and roll off a cliff, so if you see an AGENTS.md, remember to check what it specifically says before concluding it's there because the project is pro-AI
(and people who have done this have reported that writing the fuck-off message in that polite insipid tone LLMs use works better at getting them to actually stop: https://unstable.systems/@AmyZenunim/116672510693285709 )
Attached: 1 image · Content warning: slop (for a good cause)
Unsere Databroker-Recherchen ziehen weiter Kreise:
In den USA haben einige Senatoren und Kongressabgeordnete das Pentagon zu Gefahren für US-Soldat:innen durch den Handel mit Standortdaten aus der Werbeindustrie befragt: https://www.documentcloud.org/documents/28168364-ron-wydens-may-28-2026-letter-to-the-department-of-defense/
Dabei verweisen sie auch auf unsere gemeinsame Recherche von @netzpolitik_feed mit @br_data & @Wired dazu, wie sich US-Stützpunkte in Deutschland ausspionieren lassen: https://netzpolitik.org/2024/databroker-files-wie-datenhaendler-nato-und-us-militaer-blossstellen/
Digital community, we have a problem...
Christopher Kunz zeigt schmerzhaft auf, dass der implizite Deal bei #OpenSource - use but contribute - schamlos ausgenutzt wird und somit dem klassischen Problem von public goods zu erliegen droht... #AI #KI
"The dollar cost of encountering content has also fallen toward zero thanks to ad-supported platforms and massively subsidized AI agents. But the mental “decision costs” of finding accurate information have been driven skyward for consumers wandering a swamp of mostly terrible choices. The top-of-the-hour paternalism of 20th-century mass media has been traded in for the 21st-century paternalism of slop-slinging algorithms indifferent to the accuracy of the product or the compensation of journalists whose work feeds the entire ecosystem, usually without credit. What was once legible about media consumption has become increasingly illegible, depreciating our old tools of analysis and confrontation.
[...]
You probably couldn’t create a more anti-news internet if you tried (and some people seem to have tried). There are lots of things that have gone wrong for the news media in the 21st century, but the feature they have in common is the destruction of incentives to produce accurate information. Addressing these problems doesn’t require one fix but many — not just for the news outlets and journalists on the supply side, but to help out the exhausted, burned out, confused consumers on the demand side, who are getting drowned in content sludge."
https://www.niemanlab.org/2026/05/you-couldnt-create-a-more-anti-news-internet-if-you-tried/
In der Schweiz zählt ja der Schutz von digitalen Grundrechten ja nicht sehr viel... Argumente wie Steuerausfälle und Arbeitsplatzverlust umso mehr.
Eine Regulierungsfolgenabschätzung, die eine private Unternehmensplattform in Auftrag, zum revidierten VÜPF ist vernichtend:
Bis 2035 drohen jährliche Umsatzverluste von 10 Milliarden Franken, kumulierte Steuerausfälle von 7 Milliarden Franken und ein massiver Verlust von Arbeitsplätzen.
Achtung: hier geht es noch um die Vorlage VÜPF 1.0 von 2025.
Die VÜPF 2.0-Version, die ich publik machte für das Republik Magazin, ist aus rechtsstaatlicher Sicht und auch aus Perspektive des Wirtschaftsstandortes etwas besser. Trotzdem würde die Schweiz ihre Prestigeunternehmen verlieren....
Ich bin sehr gespannt ob die Regulierungsfolgenabschätzung, die vom Bund selbst beauftragt wurde, zum gleichen Schluss kommen wird...
"Chinese scholars have turned to a somewhat obscure concept to make sense of this stop-start pattern: the Collingridge dilemma, named after David Collingridge, a late professor at Aston University in Birmingham, England. In the 1980s he observed that when technologies are young, it is impossible to foresee how they will develop, so they cannot be well regulated; but as technologies spread, they can quickly reach a point where it is extremely difficult to regulate them. Although the Collingridge dilemma pops up in Western discussions every now and again, its fame is now greater in China. Since the rise of generative ai, scores of academic papers have asked whether it is possible to escape the dilemma. In March, Fudan University, one of China’s leading universities, hosted a forum devoted to the dilemma in ai governance, which put Collingridge, or “Kelingeliqi”, in news headlines."
Whether an #AI system is safe or not is a question of perspective...as the comic nicely illustrates, #alignment is a challenge that goes way beyond #technology
RE: https://social.heise.de/@heiseonline/116613061054965606
"Immerhin: Dass man vom Internet und der Digitalisierung persönlich keine Ahnung habe, damit kokettiert 2026 kein Politiker mehr."
Low bar to clear...
Lukasz Olejnik
abadidea
Ingo Dachwitz
Adfichter
Marc Ruef 
