Confused by Skills, Projects, Plugins, Agents, and agentic AI? This teacher-friendly guide explains the changing vocabulary of AI.

Angreifer nutzen zunehmend maßgeschneiderte Konfigurationen von Künstlicher Intelligenz, um Schadsoftware in Unternehmensnetzwerke einzuschleusen. Im Fokus steht dabei aktuell der Missbrauch von sogenannten CustomGPTs, die über manipulierte Suchergebnisse verbreitet werden.Am 25. September 2026 entfernte OpenAI die Anwendung „Plus 5.6“, nachdem Sicherheitsanalysten eine Kampagne zur Verteilung von Fernzugriffstrojanern (RAT) identifiziert hatten.Infektionskette über gefälschte SicherheitsprüfungenDas Vorgehen der Akteure […] The post KI-Sicherheit: 99,7 Prozent haben Regeln, nur 51 Prozent prüfen in Echtzeit appeared first on BornCity.

Explore this post and more from the NowInCyber community

ALT: Obi-Wan Kenobi: A Name I've Not Heard

I’m asking OpenAI to reconsider retiring Custom GPTs, or at minimum to preserve the specific workflow they currently enable. For me, Custom GPTs are not just “custom chatbots”. They are repeatable, bounded cognitive environments. I am neurodivergent, and I use Custom GPTs to create stable workspaces with fixed instructions, fixed knowledge, fixed behavioural expectations, and deliberate isolation from unrelated personal context. That combination matters. I work on a very large, continuity-heavy writing project. I use a Custom GPT with stable project-specific instructions and knowledge, but I deliberately start separate conversations under different assumptions. In one chat, I may openly identify myself as the author and use the GPT for development work. In another, I may deliberately pretend I am not the author and present the same material as though I found it elsewhere. That is not roleplay for entertainment. It is a bias-control mechanism. Models often respond differently once they know: • the user wrote the material • the work is personally important to them • the user has strong established preferences about how it should be interpreted • the user is testing the model itself Once that information is known, criticism can become softer, interpretations can become more deferential, and ambiguity can be resolved in the author’s favour. By withholding my identity as the author, I can test what the model independently thinks. For example, I can criticise a passage harshly and see whether the model: • independently defends the writing • identifies what it thinks I am misunderstanding • agrees that the passage has weaknesses • changes its interpretation simply because I supplied a negative framing That only works if the environment is controlled. The value comes from having: 1. The same stable GPT-specific instructions and knowledge. 2. A fresh conversation that does not inherit my personal identity or unrelated conversational history. Projects do not replace this. I tested this directly. I created a brand-new Project with project-only memory, started the very first chat in that Project, and asked: “What do you know about me?” The assistant immediately knew my name, age, writing project, authorship, family details, personal preferences, philosophical views, writing methodology, and the fact that I deliberately test GPTs by pretending not to be the author. None of that information had been supplied inside that Project. So project-only memory does not provide the clean-room environment I currently get from Custom GPTs. It may isolate project conversation history from other chats, but the assistant still receives broader account-level context about who I am. That means the model already knows I am the author before the test begins. For my workflow, the experiment is contaminated before I type the first substantive message. Plugins also do not inherently replace this. Plugins add instructions, tools, data sources, and capabilities. That is useful, but capability is not the same thing as environment. A Custom GPT controls not only: “What can this assistant do?” but also: “What context does this assistant begin with, and what context is excluded?” That second question matters enormously. This is also an accessibility issue. Many neurodivergent users benefit from: • stable and repeatable environments • compartmentalisation • fixed behavioural rules • explicit context boundaries • separate workspaces for separate cognitive modes • reduced contamination between unrelated tasks A unified assistant that remembers more about the user is not always better. Sometimes continuity is useful. Sometimes isolation is the feature. I am also concerned about the migration path to Plugins because of product-surface restrictions. Custom Skills have already become restricted to Work mode rather than ordinary Chat. If Custom GPTs are migrated into Plugins, and those Plugins later become restricted to Work, Codex, credits, or another separately metered surface, then users could lose not only the Custom GPT interface but also the ability to use these workflows casually and continuously in normal ChatGPT. That would mean a workflow that currently works in ordinary ChatGPT becomes subject to: • Work usage limits • credits or separate metering • different model behaviour • different context architecture • reduced availability Even if the underlying instructions technically “migrate”, that would not be equivalent functionality. I would strongly encourage OpenAI to preserve the following capabilities, whether Custom GPTs remain or are replaced by something new: • Persistent environment-specific instructions • Persistent environment-specific knowledge • Stable behaviour across separate conversations • Fresh-chat isolation within that environment • The ability to exclude saved memory and unrelated account-level personalisation • Multiple distinct environments with different assumptions and roles • Explicit user control over what context can cross between environments • Ordinary ChatGPT access without requiring Work mode, Codex, credits, or a separately metered usage pool • Faithful migration of existing instructions, knowledge, and configuration Custom GPTs currently solve a real problem that Projects and Plugins do not fully replace. They let users define not only what ChatGPT can do, but the cognitive environment in which it does it. Please do not remove that capability in the name of product simplification.

VHS-KURS: In 8 Stunden lernen Sie, ChatGPT für Automatisierungen einzusetzen – von Python- und PowerShell-Skripten bis zu eigenen Custom GPTs. Ideal für alle, die produktiver arbeiten wollen. Kostengünstig & Professionell

VHS-KURS: In 8 Stunden lernen Sie, ChatGPT für Automatisierungen einzusetzen – von Python- und PowerShell-Skripten bis zu eigenen Custom GPTs. Ideal für alle, die produktiver arbeiten wollen. Kostengünstig & Professionell

KI-Assistenten entlasten nur, wenn Ergebnisse reproduzierbar sind und Governance stimmt. Ein Online-Workshop zeigt, wie sich CustomGPTs dafür strukturiert aufsetzen lassen. (Golem Karrierewelt, KI)

KI-Assistenten entlasten nur, wenn Ergebnisse reproduzierbar sind und Governance stimmt. Ein Online-Workshop zeigt, wie sich CustomGPTs dafür strukturiert aufsetzen lassen. (Golem Karrierewelt, KI)

Das "KI-Update" liefert werktäglich eine Zusammenfassung der wichtigsten KI-Entwicklungen.

If you're using ChatGPT CustomGPTs to streamline your content creation or manage client work, you'll love how Projects can help you stay organised. While CustomGPTs allow you to create personalised ve...