iiinova — where your digital self keeps evolving

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 17/09/2026

And search results outlive takedowns: 47 records still returned as hits more than a month after major site seizures. Ben alerted all depicted MPs individually and provided removal guidance. 4/5

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 17/09/2026

Key finding: legislation alone doesn't protect. Countries with deepfake-specific laws did not show lower exposure. 3/5

101

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 17/09/2026

Ben scanned 5,872 nationally elected EU MPs across 160 domains hosting deepfake abuse. 138 women from 22 countries were found. Women MPs are 33x more likely to be targeted, after adjusting for age, prominence, political lean and legal context. 2/5

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 25/08/2026

Why it matters: the public manifesto alone would have pointed us the wrong way. Reading it against the forum posts, the ones nobody outside that world sees, is what let our researchers find where the worldview actually started. 6/7

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 25/08/2026

But the ingredients weren't equal. Strip it back and one grievance anchors the rest: gendered, incel-rooted resentment. Only after repeated romantic rejection does his hatred widen from specific groups to a declared misanthropy. 5/7

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 25/08/2026

What emerges is ideological bricolage. Not one doctrine but many, blackpilled misogyny, antisemitic conspiracy, anti-Muslim hatred and niche grievances, welded into a generalised misanthropy. It's a signature of Nihilistic Violent Extremism. 4/7

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 25/08/2026

So we read the two sources against each other, and they don't fully agree. Written for the world, the manifesto rejects the incel label. Written for the forum, the posts debate incel narratives openly, even solutions to female "hypergamy" (screenshot recreated from archived data). 3/7

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 25/08/2026

The manifesto was a mess of targets. Incel? Antisemite? Anti-Muslim? Each label fit a little, none fit fully. His motivation resisted easy classification, and that puzzle is where our researchers started. 2/7

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 25/08/2026

When a 16-year-old attacked his former school in Schongau on 8 July, the 19-page manifesto made headlines. What didn't: 135 posts he left on an incel forum. The account is gone now, but we archived it first. (CW: extremism, misogyny, violence) 1/7

131

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

The result is a clear compliance-utility trade-off: the more an image is decoupled from the original to raise compliance, the more classification accuracy and semantic similarity fall. Compliance has a measurable price. 8/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

As a proof of concept, we tested CAP on 400 memes, measuring how much task performance survives each transformation, alongside pixel-level and meaning-level similarity to the original image. 7/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

CAP offers an escalation ladder of three modes: inpaint just the sensitive parts, regenerate the whole image while keeping its meaning, or synthesise a brand-new image from only the abstract concept. More compliance, less resemblance to the original. 6/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

Our move: handle compliance before training. We build a Compliance-Aware Data Pipeline (CAP) that detects, transforms and evaluates sensitive data in one modular, multi-agent workflow, grounded in seven design requirements from a structured literature review. 5/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

The field has long known a privacy-utility trade-off: the more you protect personal data, the more model performance you tend to lose. We generalise it to a compliance-utility trade-off, one that spans privacy and copyright together, not privacy alone. 4/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

Today's fixes tend to treat privacy and copyright as separate problems, and mostly kick in after a model is already deployed. At web scale, manual cleanup and after-the-fact patches simply don't hold up. 3/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

It's not just music. AI is trained on billions of web-scraped images and texts too, where one item can be sensitive twice over: a photo of an identifiable person that is also under copyright. Blur the face, and the copyright problem is still there. 2/9

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 10/08/2026

On 31 July 2026 a Munich court ruled that Suno had trained its music AI on protected songs like "Atemlos durch die Nacht" without a licence, storing them in the model. Our ECIS 2026 short paper asks: what if you caught that before training? By @riekers.bsky.social & @risius.bsky.social 1/9

152

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

The key move: these aren't fixed states, they're transitions. Independence slides into dependency when AI replaces thinking. Learning turns into skill loss when use stays shallow. Productivity quietly costs connection. The model maps how one becomes the other. 7/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

The paper's new concept is competence authenticity: the felt alignment between a result and your own mastery and effort. Output can be excellent yet feel unearned. As one participant states about AI-built formulas: "If it works, it works." But is it yours? 6/8

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

And each cuts both ways. AI frees time and supports self-directed learning, but overreliance can erode the ability to work without it. It bridges language barriers in teams, yet thins out the informal peer contact that builds belonging. 5/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

Six dimensions emerged, two per need. -Autonomy: independence vs dependency. -Competence: competence shifts vs competence authenticity. -Relatedness: interaction enhancement vs alienation. Each need has an upside and a downside. 4/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

The evidence comes from five online focus groups with knowledge workers in IT, finance and marketing, analysed with an inductive grounded-theory approach. No abstract opinions about AI, but how AI-changed work actually feels day to day. 3/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

Self-determination theory says human motivation rests on three basic needs: autonomy, competence, and relatedness. The question the paper asks is simple: when AI reshapes your tasks, does it feed those needs or starve them? 2/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 28/07/2026

"You don't have to use what the AI suggests," quoting a knowledge worker in a new ECIS 2026 study, describing the freedom to take or leave what AI offers. But that freedom is only half the story. By Adeline Frenzel-Piasentin, Nikola Finze & Marie-Sophie Kosok. 1/8

111

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 22/07/2026

At DSRC, we study digital tech abuse, online communities, and the dynamics through which platform cultures can shape beliefs, behavior, and harm. Watch the full segment (in German) in the "ARD Mediathek": ardmediathek.de/video/tagest... #Looksmaxxing #TrustAndSafety #HNU 5/5

ardmediathek.de

tagesthemen 21:45 Uhr

000

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 22/07/2026

In this worldview, appearance becomes the default explanation for nearly every setback. That can reinforce harmful assumptions about gender, success, status, and social recognition — especially in online communities where these narratives are repeated and intensified. 4/5

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 22/07/2026

As Christopher explains in the segment: “The danger with looksmaxxing is that it offers an explanation for almost everything.” Not just dating frustration, but also bad grades, professional setbacks, and perceived social failure. 3/5

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 22/07/2026

Looksmaxxing describes the drive to “maximize” one’s appearance — from skincare and fitness to more drastic procedures. The issue is not necessarily the practices themselves, but the broader narratives that can emerge around them. 2/5

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 22/07/2026

Looksmaxxing is often framed as self-improvement. But when does optimizing one’s appearance become a worldview that explains almost everything? Our colleague Christopher David discussed this in a recent #Tagesthemen segment on ARD, Germany’s largest public broadcaster. 1/5

154

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

The broader takeaway: The article argues that extremist content on social media does not only spread in echo chambers. It also circulates in what the authors call hybridised spaces - where humour, mundane trends, borderline content, and outright extremism blend together. 10/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

4) Weaponised irony One of the article’s key points is how absurd situations, memes, and pop-cultural references can package extremist narratives. The authors argue that this can increase the chances of algorithmic amplification while also providing plausible deniability. 9/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

The article also points to targeted appeals. In particular, Agartha videos commonly included “looksmaxxing” and extreme self-optimisation hashtags, which the authors describe as a way of pushing videos toward audiences seen as susceptible, such as insecure young men. 8/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

3) Building network bridges Agartha content does not remain isolated. The article shows that it connects to mainstream spaces through mundane, high-traffic hashtags such as: - #roblox - #loganpaul - #gymtok 7/10

310

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

2) Dog whistles and split-second provocations The article highlights recurring coded references, including symbols and numbers, as well as extremist markers that may appear only for split seconds. This can communicate to in-groups while also testing platform boundaries. 6/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

1) Aesthetic camouflage According to the article, far-right actors use generative AI together with fantasy and science-fiction tropes to mask racial ideology behind visually engaging content. The article describes this as a way of making such ideas appear more palatable. 5/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

The article draws on: - a network analysis of more than 43,000 Agartha-related TikTok videos - a closer analysis of selected examples From this, the authors identify four mechanisms through which radical narratives are pushed to broader audiences. 4/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

The article argues that literal belief in hollow-Earth myths or Nazi UFOs is not central. Instead, “Agartha” functions as an aesthetic that can host coded and explicit far-right messages - often fused with pop-cultural references like memes, music, and AI-generated imagery. 3/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

A new article by our colleagues @cvdavid.bsky.social and @risius.bsky.social examines the "Agartha" trend and how white supremacist narratives can be embedded in absurd, aesthetic, and meme-driven content. 2/10

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 13/05/2026

How does extremist ideology travel into mainstream TikTok feeds - without looking like classic extremist propaganda? 👀 A Thread 🧵 1/10

164

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 24/04/2026

The study highlights an important design tension for AI in search: How can systems support reflection, verification, and better judgment — without creating too much friction. 7/8

110

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 24/04/2026

A key reason was a trade-off: More time spent chatting with the copilot often meant less time exploring actual search results. 6/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 24/04/2026

Users engaged deeply with the copilot and showed signs of metacognitive reflection. But answer correctness and search engagement did not significantly improve overall. 5/8

100

Digital Security Research Collaboratory Neu-Ulm @dsrc.bsky.social · 24/04/2026

In a pre-registered randomized controlled trial (N=261), three conditions were compared: • 10 blue links • AI overview • AI overview + copilot 4/8

100