
Professor @uni_bamberg_of | privacy, information security, and ethics | digital teaching @dikule_uba | enthusiastic speaker | typography nerd | EN/DE
Bamberg, DE
Joined March 2008
- Tweets9.4K
- Following277
- Followers770
- Likes1.9K
Your regular Chrome already auto-updated to 156, so you think you're safe
You're not
The real danger here is Playwright, Puppeteer, and MCP browser tools running inside Cursor, Claude Code, or Terminal
1. Your IDE / Terminal has macOS Accessibility permissions
2. It spawns a cached Playwright Chromium (usually stuck on v131โ151 or whatever)
3. That browser inherits your IDE's Accessibility permission
4. Any site your agent visits can use the GPU process to press Cmd + Space, open Terminal, and run commands on your Mac ๐
Check ~/Library/Caches/ms-playwright and ~/.cache/puppeteer right now and update to Playwright 1.64.1+ (Chromium 156)
๐จ BEWARE: Chrome Sandbox Escape
Full remote command execution + full chain capabilities
October has been wild
pwn.ai/blog/chrome-escape
Dominik Herrmann retweeted
Initially, I had Claude make this animation in 2D. It wrote the piano score and drew every frame in JavaScript.
I gave this 2D version as reference to another Claude. Because the whole film was code, Claude was able to use its exact choreography and make the 3D one in Blender.
Dominik Herrmann retweeted
Over the last few weeks I've been running a few workshops on Practical AI for academics. I thought I would share the slides and reflect a bit on what I have learned. You can find the slides and some material here: claesbackman.com/ai_workshopโฆ
This AGENTS.md fixes the ๐๐ต๐ฟ๐ฒ๐ฒ ๐บ๐ผ๐๐ ๐ฎ๐ป๐ป๐ผ๐๐ถ๐ป๐ด ๐ต๐ฎ๐ฏ๐ถ๐๐ of Codex and Claude Code: writing backward-compatible code for a project nobody uses yet, never deleting old code and docs it no longer needs, and adding a pile of tests just to change a button color.
After using them every day for months, the lines I'd want in there from day one are the ones at the top: if the project has no real users yet, skip compatibility and migrations, and go straight to ๐๐ต๐ฒ ๐๐ถ๐บ๐ฝ๐น๐ฒ๐๐ ๐ฐ๐ผ๐ฑ๐ฒ ๐๐ต๐ฎ๐ ๐๐ผ๐ฟ๐ธ๐.
Then a four-step loop: docs first โ tests โ smallest implementation โ clean up.
Copy and paste it into your AGENTS.md ๐
A week of Claude Code sessions is too much to remember. Every weekend, have Opus 5.5 vibe-code a quick ๐๐ฒ๐ฒ๐ธ๐น๐ ๐ฐ๐ฎ๐น๐ฒ๐ป๐ฑ๐ฎ๐ฟ like this from your local session logs.
Then give it ๐ฎ ๐๐๐ฏ๐ฎ๐ด๐ฒ๐ป๐ ๐๐ต๐ฎ๐ ๐ผ๐ป๐น๐ ๐ฏ๐๐ถ๐น๐ฑ๐ ๐๐ฒ๐ฒ๐ธ๐น๐ ๐ฐ๐ฎ๐น๐ฒ๐ป๐ฑ๐ฎ๐ฟ๐:
โ Name it week-calendar, set its effort to medium, and preload a design skill
โ The first time, it asks which colors you like and whether to color by project or by task, and saves that to memory. After that, every calendar follows it
โ Call it when you want to look back on the week or write a weekly report. It regenerates the calendar and writes up a 3-line weekly report for you
โ The calendar shows four things: when you worked each day, what each session did, ๐๐ต๐ถ๐ฐ๐ต ๐๐ฒ๐๐๐ถ๐ผ๐ป๐ ๐๐๐ผ๐ฝ๐ฝ๐ฒ๐ฑ ๐๐ถ๐๐ต๐ผ๐๐ ๐ฎ ๐๐ถ๐ป๐ด๐น๐ฒ ๐ฐ๐ผ๐บ๐บ๐ถ๐, and how many hours each project took this week
Send this prompt to Claude Code ๐
"Set up a subagent that only builds weekly calendars:
1. Create week-calendar in ~/.claude/agents: model: opus, effort: medium, memory: user. Preload the design skills I have installed (for example impeccable). Give it Bash so it can run git log and the script that builds the calendar. It only writes files to ~/.calendar/ and its own memory.
2. The first time, it asks me what style I like: dark or light, one accent color, color by project or by task type, and whether the week starts on Monday or Sunday. It saves my answers to memory and follows them every time.
3. The calendar is one HTML file. The script writes the data straight into it, so it opens with a double-click. Use this week's session logs: Claude Code's are in ~/.claude/projects/ (skip the subagents folders), and if I use Codex, add ~/.codex/sessions/. Before writing the script, open one file to check the fields. Convert times to my local time zone, and use the cwd in each record to tell projects apart. One block per session; start a new block after any gap longer than 30 minutes. Clicking a block shows the project, start and end time, the model used, the session title (if there isn't one, sum up my first typed message in 3 to 5 words), my first typed message, and my commits and changed files in that repo during that time (skip folders that aren't git repos). Sessions that stopped without a single commit get a red outline and also go in a separate list. At the top, total this week's hours per project, counting parallel sessions only once. Keep everything local.
4. Add a rule to ~/.claude/CLAUDE.md: when I say "what did I do this week" or "write my weekly report", have week-calendar regenerate the calendar first, then write a 3-line weekly report from it: what I mainly got done this week, which project took the most time, and which no-commit sessions to pick up first next week.
Show me the contents of the files you'll create or change, then explain the whole flow in words a 10-year-old could follow. Don't write anything until I confirm."
I asked Opus 5.5 to write, produce, and create me an educational 'cartoon' about The History of Light - using whatever resources it had available to it.
I didn't expect it to make a 36 minute video and spend $200+ dollars though ๐
The study of light is genuinely fascinating, and we have spent thousands of years trying to understand it. This is that story.
Opus 5.5 wrote and produced the entire film. I told it to use any keys in my keychain to generate the voices, create any images or video it might need, and to create whatever motion graphics it could to explain concepts in electromagnetism.
It used
> Gemini 3.8 Flash TTS for the narration, and I must say, it did an excellent job.
> Lyrica 3 for the background music, which is also essentially perfect.
> GPT Image 2.5, Nano Banana Pro and NB2 Flash for some of the scenes imagery
It's certainly not perfect but I was still really impressed with how well it did producing the film. It's very interesting if you're a physics nerd like me.
Opus is genuinely creative in a way that I've not really seen with other models. I just gave the model a few constraints and just let it make all the decisions end to end.
It even wrote itself credits at the end lol.
It's going to be so fun to create bespoke educational videos like this for my children to learn from
Dominik Herrmann retweeted
๐ Days 3 & 4 of #ARES2026 featured research across AI, critical infrastructure, cyber risk, forensics, PQC & more.
๐ Congrats to Best Paper winners Ralf Gundelach, Michael Mรผhlhauser & Dominik Herrmann!
โ๏ธ Day 3 ended at Flygvapenmuseum with the banquet and choir performance. โ at Sweden
Dominik Herrmann retweeted
Let me blow your mind real quick:
When you use Remote Desktop (RDP), Windows secretly takes screenshots of what you are doing.
Itโs called the RDP Bitmap Cache.
To make the connection faster, Windows saves small tiles (images) of the remote screen to your hard drive in a bin file.
Even if the session is over and the remote server is destroyed... your laptop still holds the cache files.
Forensics teams use tools like BMCViewer to stitch those tiles back together.
They won't just see logs but the literal email, document, or picture you were looking at.
๐
Dominik Herrmann retweeted
In security, when you do your job perfectly, nothing happens. And people don't see when nothing happens.
Dominik Herrmann retweeted
You know that feeling when your advisor says 'but what's the gap?' & your mind goes blank Here's your rescue!
Learn your Gap types & how to address them โคต๏ธ
โ Theoretical: when theories fail to explain reality
โ Methodological: when current methods fall short
โ Empirical: when we lack evidence to explain something
โ Conceptual: when key definitions are muddy
โ Temporal: when research hasn't kept up with changes
โ Spatial: when studies ignore geographic differences
โ Literature: when studies fail to build on existing knowledge
Each has distinct ways to address them.
What most miss?
Connecting MULTIPLE gap types creates the strongest research justification.
๐ฌ Reweet & ๐๐ฐ๐ฎ๐ฎ๐ฆ๐ฏ๐ต ๐น๐๐๐๐๐๐๐ ๐ฎ๐๐ ๐ท๐ซ๐ญ ๐ช๐ง ๐บ๐ฐ๐ถ ๐ธ๐ข๐ฏ๐ต ๐ต๐ฉ๐ฆ ๐ณ๐ฆ๐ด๐ฆ๐ข๐ณ๐ค๐ฉ ๐จ๐ข๐ฑ ๐ธ๐ณ๐ช๐ต๐ช๐ฏ๐จ ๐ต๐ฆ๐ฎ๐ฑ๐ญ๐ข๐ต๐ฆ๐ด
------------------------------
Follow @RaziaAliani to get more useful AI in Research content (no clickbait stuff!) in your feed.
& join 16K+ researchers & get FREE exclusive tips on using AI in research
๐ in bio
Dominik Herrmann retweeted
My workflow for AMAZING creative writing with LLMs:
The trick is to give the LLM as much opinionated context as possible, so the writing is deeply influenced by your unique ideas.
To do this, we guide the LLM into a Q&A mode before writing.
Hereโs how:
1. Pick a topic and format.
2. Prompt GPT-4.5 with something like:
โIโm looking to write about [topic], in [format] format.
I want you to start by asking me 5-10 questions to understand my perspective on this topic. Iโll answer with long-form text (Iโll use transcription so you get tons of context). Then, based on those answers, ask another round of questions to dig deeper and flesh out perspectives.
Repeat this process until I tell you to start writing. At that point, take everything I have shared, and create an outline. Run it by me. Once I approve it, create my unique piece of writing.โ
3. Use voice-to-text to answer the questions GPT-4.5 asks. Repeat as many times as youโd like until you feel the model has enough context to do well. Then, tell it to outline, check it out, and if you like it, tell it to write.
Dominik Herrmann retweeted
90% of academic papers I read are now AI-assisted.
Most researchers are in complete denial.
I'm a professor who's been brutally using AI for 18 months.
Here's what I learned that could save your career:
Dominik Herrmann retweeted
Yes.
Writing is not a second thing that happens after thinking. The act of writing is an act of thinking. Writing *is* thinking.
Students, academics, and anyone else who outsources their writing to LLMs will find their screens full of words and their minds emptied of thought.
Dominik Herrmann retweeted
I vibe coded an app to chat with my Kindle books. (To make a book chattable it flips through the pages using an automated browser and screenshots / OCRs them.) I've been using it for a couple of weeks and it's become an important part of my learning and research workflows.
It is NOT a replacement for actually reading books. There is simply no alternative to spending a few hours with a book to really process the ideas. Besides, I enjoy reading too much to want to take shortcuts.
Here are a few things the chat interface lets me do:
- Create summaries that I edit and save in a personalized knowledge wiki (that I also vibe coded) so that I can quickly refer back to books I've read.
- Help me decide which books to read in the first place (I read a lot of books but there are 10x as many books I don't have time to get to, so I have to be selective).
- If a book mostly tells me stuff I already know, quickly zoom into the parts that are new (in the past I would have read the first chapter or two, realize that I didn't learn much, and put the book down).
- If I engage with a book critically in my own writing, do a sanity check by asking questions like "did the author address this counterargument anywhere?" (This is the least frequent use case but perhaps the most important one because the peace of mind I get from it is a big deal.)
Dominik Herrmann retweeted
๐ข People that I know often send me their draft papers to get feedback. I enjoy doing this sort of informal peer review and would like to open it up to everyone. Details in the Q&A below.
What is informal peer review and why is it helpful?
The *formal* peer review system serves to gatekeep publication slots in prestigious journals, conferences, and other publication venues. Unfortunately, providing useful feedback takes a back seat. Informal peer review fills this gap. It is unstructured, voluntary, and private. It helps improve papers before they are submitted for formal peer review or released online as preprints.
Why are you doing this?
- Iโm just paying it forward. Almost all of my own writing benefits from informal peer review.
- Itโs impactful โ as a senior researcher I can often provide a useful complementary perspective with a quick look at a paper.
- Itโs gratifying, and Iโm at a career stage where I can just do stuff that I find meaningful without worrying about whether it will contribute to my CV.
How should I send you my paper?
Just email it to me at [email protected]. Ideally include both a PDF and a source document (Google doc, Word, Overleaf, etc.)
Will you give feedback on all submissions?
It entirely depends on what my workload is like and how many requests I get. I tend to have more time in the summer than during the semesters. For now, I plan to do 2-3 papers per week. I will try to get back to you within a couple of weeks (much quicker than formal peer review).
What topics are in scope?
Take a look at my writing to get an idea of what my areas of interest/expertise are. Iโm very much open to papers outside my discipline(s) but where my perspective can be helpful.
How should I acknowledge your feedback?
List me in the acknowledgements section of your paper. In addition, I might maintain a public list of titles/authors of papers Iโve given feedback on. (I havenโt decided yet.)
Does it have to be papers or can it also be ... ?
It can be blog posts, essays, policy papers, etc. It doesnโt necessarily have to be scholarly work either. But nothing commercial unless youโre paying me.
===
If you have any other questions, feel free to ask. This is all a giant experiment. It might work really well or it might not work at all. I will periodically post updates on how itโs going.
Dominik Herrmann retweeted
Most researchers make a fatal mistake in their proposal's introduction, but understanding the "Why-What" sequence can change everything...
Here's a 15-part structure I use that makes it simple.
Let's break it down into 7 broad steps:
Dominik Herrmann retweeted
A hypothesis on the accelerating decline of reading:
* Broadly speaking, people read for pleasure/entertainment and for learning/obtaining information.
* Reading for pleasure has been declining for a while and is being replaced by videos (very sharply among young people). This trend will surely continue.
* Reading for obtaining information is getting intermediated by chatbots. We are in the very early stages of this shift, so I think people underappreciate the magnitude of what's coming. It's not just that AI replacing traditional web search. Even when it comes to reading news articles, business documents, or scientific papers, the vision that tech companies are pushing on us is AI summarization + synthesis + Q&A.
* We don't have to accept this, but I predict that most people will. It's a tradeoff between speed/convenience and accuracy/depth of understanding โ the same tradeoff that was once offered to us when it became possible to search the web to look up a quick fact as opposed to reading about the topic in depth in an encyclopedia.
* Just as most people in most cases prefer a shallow web search over deeper reading, most people in most cases will prefer AI-intermediated access to knowledge. Traditional reading won't disappear, but people will do it vastly less often, except in hobbyist reading communities and professions where traditional reading is needed.
* The decline of reading-for-pleasure (due to video) and reading-for-information (due to AI) will accelerate each other, as reading text without an intermediary will come to be seen as a chore.
* Personally, I find this sad. But while it's tempting to moralize all this, I think that's unproductive. Yelling at individuals to resist new media has been done for centuries and has never worked.
* Even if people individually rationally choose these tradeoffs, I think we collectively lose something; critical reading skills are arguably essential for a democracy. We need to figure out what to do about that.
Dominik Herrmann retweeted
People who say they're "bad at tests" tend to just not be very smart.
When they blame their low performance on anxiety instead of low intelligence, that's generally just a poor excuse.
Dominik Herrmann retweeted
I just realized something most people are going to lose when (as they inevitably will) they start using AIs to write everything for them. They'll lose the knowledge of how writing is constructed.
Dominik Herrmann retweeted
creating a mcp is actually superrr easy
3 steps process to make an mcp in 10 minutes:
- copy this file modelcontextprotocol.io/llmsโฆ
- paste in cursor
- write a detailed prompt on what you want the mcp to do
and done! bookmark this tweet





























