DEV Community: Alexander Loth

From Misinformation to Agentic AI: Where My Research Is Heading

Alexander Loth — Tue, 31 Mar 2026 13:05:18 +0000

Two years ago, I started studying AI-generated misinformation. I wanted to understand how large language models produce convincing false content, how quickly that content spreads, and whether humans can even tell the difference anymore. That work led to four papers at The Web Conference 2026, tools like JudgeGPT and RogueGPT, and a growing concern I could not shake: the problem is bigger than text.

Because while I was studying how AI breaks trust through information, something else was happening. AI was gaining the ability to act.

The Shift I Noticed

Misinformation research forced me to think carefully about trust. Can you trust that a piece of content is what it claims to be? Did a human write it? Is the source legitimate? Can verification tools keep pace with generation quality?

These are information integrity questions. But they turn out to be a specific instance of a broader challenge: how do you maintain trust and oversight when AI systems operate with increasing autonomy?

The same questions that apply to AI-generated text apply -- with more urgency -- to AI agents taking actions on your behalf. When an agent sends an email, edits a document, or executes a command in your environment, how do you know it did what you intended?

Misinformation is a trust problem in the information layer. Agentic AI is a trust problem in the action layer.

PowerSkills as a Practical Case Study

One project I have been building is PowerSkills -- Windows automation skills for AI agents. It gives agents structured access to Outlook (email and calendar), Edge browser via Chrome DevTools Protocol, desktop automation, and shell commands.

PowerSkills is open source (MIT license) and installable via AgentSkills:

npx skills add aloth/PowerSkills

Every command returns a consistent JSON envelope:

{
  "status": "success",
  "exit_code": 0,
  "data": { ... },
  "timestamp": "2026-03-06T16:00:00+01:00"
}

Building it clarified something: the agent-tool interface is a design problem, not just an engineering one. Decisions about what to expose, what to restrict, and how to structure output all affect how safely and predictably an agent operates. A well-designed tool surface makes agent behavior more auditable.

Open Questions for Builders

If you are building agentic systems, these are the questions I think matter most:

Verification of agent actions. When an agent completes a task, how does a human confirm it did the right thing?
Trust calibration. How should trust in an agent accumulate or decay based on observed behavior?
Agent-tool interface design. How do you design tool interfaces that make unsafe actions harder and correct actions clearer?
Multi-agent oversight. As agents orchestrate other agents, who watches the watcher?

What Is Next

More empirical work with PowerSkills, expanding the agent-tool interface research, and connecting the information integrity thread to agentic AI. The trust questions do not go away just because the AI got more capable. They get harder.

Full post: alexloth.com

JudgeGPT & RogueGPT: Building Open-Source Platforms for AI Misinformation Research

Alexander Loth — Tue, 17 Mar 2026 14:08:00 +0000

Can people tell AI-written news from human-written journalism? As large language models grow more capable, the answer is becoming increasingly uncomfortable. This is the question at the heart of two open-source research platforms: JudgeGPT and RogueGPT.

Both are licensed under GPLv3 and have companion papers accepted at The Web Conference 2026 (WWW '26).

The Problem: Industrialized Deception

Generative AI has created an asymmetric arms race. Producing convincing synthetic news now costs almost nothing. Detecting it reliably does not. Two papers at WWW '26 address this:

"Industrialized Deception: The Collateral Effects of LLM-Generated Misinformation on Digital Ecosystems" (arXiv:2601.21963) -- systemic effects of LLM-generated misinformation on trust networks.
"Eroding the Truth-Default: A Causal Analysis of Human Susceptibility to Foundation Model Hallucinations and Disinformation in the Wild" (arXiv:2601.22871) -- key finding: the human truth-default is being measurably eroded by LLM-generated content.

RogueGPT: Controlled Stimulus Generation

RogueGPT is a Python framework for generating controlled news stimuli. The current corpus contains 2,663 multilingual news fragments: 37 model configurations across 10 providers (OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, Microsoft, Zhipu, Moonshot, Qwen, MiniMax), 4 languages, 3 formats, 5 journalistic styles per language, and 222 human-sourced fragments as experimental anchors.

Three interfaces over a shared data layer: Streamlit app, CLI, and an MCP server exposing tools for AI agent integration.

git clone https://github.com/aloth/RogueGPT
pip install -r requirements.txt
python cli.py ingest --text "..." --model "gpt-4o" --language en --style nyt --format article
python cli.py retrieve --model "gpt-4o" --language en --limit 10

JudgeGPT: Human Evaluation at Scale

JudgeGPT is a live Streamlit platform collecting human judgments on news authenticity. Participants evaluate fragments on three 7-point scales: source attribution (human vs. machine), veracity (legitimate vs. fake), and topic familiarity.

After each submission, participants see the ground truth and the specific model that generated the content. A shareable score card is generated every 5 responses.

Live survey: judgegpt.streamlit.app

Why It Matters for Developers

Every fragment has full provenance: model, parameters, seed. This enables questions beyond can humans detect AI -- which models are hardest to detect? In which languages? By which demographic groups?

Corpus on Zenodo (academic access): DOI: 10.5281/zenodo.18703138

Both repos are GPLv3. Contributions welcome.

GitHub: aloth/JudgeGPT | aloth/RogueGPT
arXiv: 2601.21963 | 2601.22871

Trackless Links: Building a Safari Extension for Automatic URL Cleaning

Alexander Loth — Tue, 10 Mar 2026 21:44:41 +0000

Every link you click carries invisible baggage. Those cryptic strings at the end of URLs -- utm_source, fbclid, gclid, twclid -- are tracking parameters. They tell advertisers where you came from, what you clicked, and how to follow you across the web.

Most people never notice them. But once you do, they are everywhere.

The Problem

A typical tracked URL looks like this:

https://example.com/article?utm_source=twitter&utm_medium=social&utm_campaign=spring2026&fbclid=IwAR3x...

Strip the tracking, and all you need is:

https://example.com/article

Those extra parameters tell the destination site exactly which campaign brought you, which platform you came from, and sometimes which specific ad you clicked. When you share a tracked link with someone else, you pass along that tracking context and potentially link their activity to yours.

How Trackless Links Works

Trackless Links is a Safari extension for iOS and macOS that intercepts URL loads and strips tracking parameters before the request reaches the server.

The extension uses Safari's declarativeNetRequest API (content blockers) combined with a curated ruleset that matches known tracking parameter patterns:

UTM parameters (utm_source, utm_medium, utm_campaign, utm_content, utm_term)
Facebook (fbclid, fb_action_ids, fb_action_types)
Google (gclid, gclsrc, dclid)
Twitter/X (twclid)
TikTok (ttclid)
And dozens more from various ad networks and analytics platforms The key architectural decision: everything runs on-device. The extension uses Safari's native content blocking API, which means URL matching happens in Safari's process without any network calls or external dependencies.

Bonus: Domain Credibility with CRED-1

Trackless Links also ships with the CRED-1 dataset -- 2,672 domains with five independent trust signals aggregated from established credibility databases. When you visit a site with known credibility issues, the extension can flag it.

CRED-1 grew out of my doctoral research on AI-driven disinformation at Frankfurt University of Applied Sciences. The dataset is open source (CC BY 4.0) and documented in a peer-reviewed paper accepted at ACM WebConf 2026.

Try It

GitHub: github.com/aloth/trackless-links (30 stars)
App Store: Trackless Links (free, iOS) and Trackless Links Pro (iOS + macOS)
Blog post: alexloth.com/trackless-links-clean-urls-zero-tracking

If you find a tracking parameter that slips through, open an issue. PRs welcome.

PowerSkills: Giving AI Agents Control Over Windows with PowerShell

Alexander Loth — Sat, 07 Mar 2026 08:41:04 +0000

If you're building AI agents that need to interact with Windows, you've probably noticed: most agent tooling assumes Linux or macOS. Windows automation is an afterthought.

But enterprise work happens on Windows. Outlook holds the emails. Edge holds the browser sessions. PowerShell is the automation backbone.

PowerSkills bridges this gap.

What is PowerSkills?

PowerSkills is an open-source PowerShell toolkit that gives AI agents structured control over Windows - Outlook email, Edge browser, desktop windows, and system operations. Every action returns clean, parseable JSON.

The Four Skill Modules

Outlook - Read inbox, search emails, send messages, access calendar events via COM automation
Browser - Control Edge via Chrome DevTools Protocol (CDP) - list tabs, navigate, take screenshots, interact with the DOM
Desktop - Manage windows, capture screenshots, read/write clipboard, send keystrokes via Win32 API
System - Query system info, manage processes, execute commands, read environment variables

Structured JSON Output

Every action returns a consistent envelope. No more parsing free-text output:

{
  "status": "success",
  "exit_code": 0,
  "data": {
    "hostname": "WORKSTATION-01",
    "os": "Microsoft Windows 11 Pro",
    "memory_gb": 32
  },
  "timestamp": "2026-03-06T17:30:00Z"
}

Agents check status, extract data, handle errors - no regex needed.

Two Ways to Run

# Dispatcher mode
.\powerskills.ps1 system info
.\powerskills.ps1 outlook inbox --limit 5
.\powerskills.ps1 browser tabs
.\powerskills.ps1 desktop screenshot --path C:\temp\screen.png

# Standalone mode
.\skills\system\system.ps1 info
.\skills\outlook\outlook.ps1 inbox --limit 5

Agent-Friendly by Design

Each skill includes a SKILL.md file with structured metadata - name, description, available actions, and parameters. AI agents can discover and understand capabilities without hardcoded instructions.

Getting Started

No package manager, no installer. Just PowerShell 5.1+ and Windows 10/11:

Clone or download the repository
Run: .\powerskills.ps1 list
For browser skills: launch Edge with --remote-debugging-port=9222

Note: If scripts are blocked, set the execution policy: Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

Check It Out

PowerSkills is MIT licensed. Contributions, issues, and stars are welcome:

github.com/aloth/PowerSkills

If you're building agents that need to work with Windows, I'd love to hear how you're approaching the problem. What other Windows capabilities would be useful for your agent workflows?