←
Daniel Thilo Schroeder126 posts
Daniel Thilo Schroeder avatar

Daniel Thilo Schroeder

@daniel_thilo

daniel-thilo.bsky.social Computers, Graphs, and Conspiracies Curious about most things, most of the time.

◷ Joined November 2017▧ 18 Photos and videos
Gary Marcus avatar
⟳ Daniel Thilo Schroeder retweeted
Gary Marcus@GaryMarcus·Sep 19

yikes!!!! straight out of the nightmares @daniel_thilo Jonas Kunst and I and others warned about earlier this year in our Science paper (and Substack post) re swarms of AI agents generating misinformation.

Post attachment
Lukasz Olejnik avatar
⟳ Daniel Thilo Schroeder retweeted
Lukasz Olejnik@lukOlejnik·Aug 26

An off-the-shelf VM is not enough to contain a modern, cyber-capable AI agent. AI agent escaped a VM three times. First via known bugs, then via missed patches, then it found 0-days and chained them. AI can hack, and agents can persist, research, adapt, and weaponize software flaws faster than patch cycles can handle it. blog.trailofbits.com/2026/08…

Post attachment
Bjørn Ihler avatar
⟳ Daniel Thilo Schroeder retweeted
Bjørn Ihler@bjornih·Jul 8

Fifteen years ago I was on Utøya when 77 were killed. I spoke with the BBC World Service about that and the years since. Trying to prevent violence led to the work I now do at Revontulet, where we map extremist networks so they can be stopped earlier. bbc.co.uk/programmes/w3ct8jk…

War Studies avatar
⟳ Daniel Thilo Schroeder retweeted
War Studies@warstudies·Aug 13

In Dr @lukOlejnik & his co-authors’ paper, experiments showed that AI agents could autonomously change simulated attitudes across several scenarios, without human action. The research provides a framework for studying AI-enabled influence campaigns. 🔗arxiv.org/pdf/2608.10920

Post attachment
Lukasz Olejnik avatar
⟳ Daniel Thilo Schroeder retweeted
Lukasz Olejnik@lukOlejnik·Aug 12

Releasing a multi-month work today! We conducted the first fully autonomous information operation (under controlled conditions). It ran the campaigns end-to-end. It's can work in full autonomous mode. After initial human guidance it can make strategic, tactical and individual decisions, including develop narratives, establish a presence, publish and amplify material, measure progress and use it to guide later actions. No human directed those individual actions. arxiv.org/abs/2608.10920

Post attachmentPost attachment
Daniel Thilo Schroeder avatar
Daniel Thilo Schroeder@daniel_thilo·Aug 12

This makes the threat of malicious AI swarms concrete. The key shift isn’t AI-generated content—it’s autonomous, persistent coordination. Malicious AI swarms are no longer just hypothetical. We can now study their dynamics under controlled conditions.

Releasing a multi-month work today! We conducted the first fully autonomous information operation (under controlled conditions). It ran the campaigns end-to-end. It's can work in full autonomous mode. After initial human guidance it can make strategic, tactical and individual decisions, including develop narratives, establish a presence, publish and amplify material, measure progress and use it to guide later actions. No human directed those individual actions. arxiv.org/abs/2608.10920

Post attachmentPost attachment
Lukasz Olejnik avatar
⟳ Daniel Thilo Schroeder retweeted
Lukasz Olejnik@lukOlejnik·Aug 12

We conducted the first fully autonomous information operation under controlled conditions. It ran inside a simulated social platform.

Lukasz Olejnik avatar
⟳ Daniel Thilo Schroeder retweeted
Lukasz Olejnik@lukOlejnik·Aug 11

If all goes well, big news tomorrow. We're bringing a warning to societies. AI will have quite a broad effect, and it's important to understand this.

Hello Future avatar
⟳ Daniel Thilo Schroeder retweeted
Hello Future@Orange_Future·Jun 1

🤖Comment protéger le débat public face à la menaces de essaims #IA ? Les chercheurs @daniel_thilo et Jonas R. Kunst proposent une stratégie mêlant IA de protection pour les citoyens, observatoires mondiaux et régulation des mécanismes économiques de la désinformation. 👇oran.ge/4bL3vnM

0:18
Florian Gallwitz avatar
⟳ Daniel Thilo Schroeder retweeted
Florian Gallwitz@FlorianGallwitz·Feb 22

KI-Agenten ein Stück Software programmieren zu lassen ist eine Sache, aber Openclaw (mit Codex) eine ganze Reihe von Machine-Learning-Experimenten durchziehen zu lassen und nach einer halben Stunde über iMessage die Ergebnisse zu bekommen, fühlt sich noch ein wenig surrealer an.

Jay Van Bavel, PhD avatar
⟳ Daniel Thilo Schroeder retweeted
Jay Van Bavel, PhD@jayvanbavel·Feb 13

The distinction between genuine grassroots activism and automated influence operations iscollapsing. While policy debates focus on bot farms, in a new paper we describe a distinct threat to democracy emerging via partisan coordination apps and artificial intelligence–what we term “cyborg propaganda" jonaskunst.net/wp-content/up… Led by Jonas Kunst

Post attachment