Tak. God pointe: bsky.app/profile/solm...

(Vi har jo alle gang i forskellige podcasts og projekter, men det ville susme være hyggeligt!)

Vi samler bandet igen!

Mange tak. Ærlig talt lidt på røven og, ja, rørt, over den kommentar. Dejlig slutning på en hård uge.

Haha. Det går nok ;)

Det må være ved mundingen på underboring - ingen kabler ligger på bunden helt tæt på land

Husker det som i går - han havde et lille transportabelt fjernsyn med til topmøde og roligan halstørklæde på.

Arbejder uden manus så nogle gange siger jeg forkerte ting - del af konceptet. Marcel stiller mig spørgsmålene og så generer jeg løs :) vi prøver at fange evt faktafejl i klip og det er indtil videre gået i godt to år uden store svipsere. Har en dag til red møder, research og optagelse alt i alt

Ja! Det var nemlig det han sagde. Tak!

When 20-year-old hacker tricks can make Google's AI fabricate DR exposés, it's a warning: if truth is to stand a chance, we need editorial, human-made journalism like this gem by Jakob Ussing & team (w assist from me). Main story: www.dr.dk/nyheder/peng... news analysis: www.dr.dk/nyheder/peng...

Who's behind this? Flemming Splidsboel of the Danish Institute for International Studies: "Russia is intensely focused on manipulating chatbots—poisoning them to generate answers they want. It's obvious Russians could be behind this." The Russian embassy Copenhagen did not respond.

As @garymarcus.bsky.social has long argued: LLMs fundamentally don't know what's true or false. Hallucinations aren't a bug, they're a feature. Google has bet its business on technology that guesses what is most likely true or false for you, instead of sending you off to find truth for yourself.

A recent EBU/BBC study www.bbc.co.uk/mediacentre/... found 45% of AI assistant answers had issues. 20% contained factual errors or hallucinations. Google's Gemini performed worse, with "significant issues in 76% of responses, more than double the other assistants, largely due to its poor sourcing"

After we confronted Google, the fake summaries stopped appearing in simple searches. Their response was that AI summaries are "99% spam-free" and that they have "clear policies" against this. Not exactly addressing our questions or the vulnerability, which seems fundamental.

Jan Kaastrup (CSIS) told us he used this exact trick 20 years ago as a pentester. Back then, it tricked individual websites. Now it can poison AI Overviews that serve 8.5 billion searches per day: "It seems to easy to do. I think Google has a clear share of the responsibility"

The technique is both elegant and old: append fake headlines to URLs on sites that don't reject malformed requests. Google's crawler indexes them, et voilà: Gemini summarizes "news" that doesn't exist. We found many such URLs across multiple countries with the same poisoned slug.

If you click the links, you don't find the fake news story. Here's one: www.kaa.go.ke/corporate1/a... <- note the affixed "?s-news" string sending you to the airport page? The "news" exists only in thattheURL but ends up in Google's AI summary. These highly ranked "news" pages were not compromised.

In October, my colleagues at DR exposed how 1,900+ tons of sanctioned Flügger paint was shipped to Russia via Estonia. Days later, a counter-narrative appeared on Google: "New investigation reveals Flügger paint was never sent to Russia", citing Kenya Airports Authority and a Hungarian university(!)

[the attackers] "made Google believe that this is credible news (..) by fabricating some fake URLs pointing to credible, legitimate websites" said Jan Kaastrup of CSIS. Jens Monrad (ex-Mandiant): "Simply using poorly configured websites, they succeeded in making Google's AI run with a false story"