charemma@security:~$
← back to the homepage Knowledge Topic overview

AI security

All knowledge articles on risks in LLM applications and AI pipelines. Each entry explains the topic briefly and links to the in-depth article.

LLM applications

Risks in applications built on language models 10 topics
  • KnowledgeLLM01

    Prompt injection

    Inputs – typed directly or hidden in documents, web pages and emails – make a model ignore its instructions or follow someone else’s.

    Article coming soon
  • KnowledgeLLM02

    Sensitive information disclosure

    A model reveals confidential data – from its training data, the conversation context or connected systems.

    Article coming soon
  • KnowledgeLLM03

    Excessive agency

    An agent can do more than its task requires: too many tools, too broad permissions, too little human confirmation.

    Article coming soon
  • KnowledgeLLM04

    Supply chain

    Risks that enter an application through third-party models, datasets, plugins and libraries.

    Article coming soon
  • KnowledgeLLM05

    Data and model poisoning

    Manipulated training, fine-tuning or embedding data deliberately change how a model behaves.

    Article coming soon
  • KnowledgeLLM06

    Unbounded consumption

    Unthrottled requests drive up cost and load, or allow a model to be systematically extracted.

    Article coming soon
  • KnowledgeLLM07

    Misinformation

    Convincingly worded but wrong output that people or downstream systems rely on.

    Article coming soon
  • KnowledgeLLM08

    Hidden context exposure

    Hidden instructions and context – such as system prompts – become visible and reveal logic, rules or secrets.

    Article coming soon
  • KnowledgeLLM09

    Vector and embedding weaknesses

    Weaknesses in RAG systems: missing access separation, manipulated or extracted vector data.

    Article coming soon
  • KnowledgeLLM10

    Improper output handling

    Model output is processed without validation – as HTML, SQL or a shell command – and reopens classic vulnerabilities.

    Article coming soon

AI pipelines

From the data to the inference endpoint 5 topics
  • KnowledgePIPE·01

    Data & RAG ingestion

    Whoever can feed training, fine-tuning or retrieval data can influence how the system behaves.

    Article coming soon
  • KnowledgePIPE·02

    Model supply chain

    Provenance and integrity of models, weights and dependencies – including registries and unsafe file formats.

    Article coming soon
  • KnowledgePIPE·03

    Agents & tool access

    MCP servers, tool permissions and automation decide what a manipulated agent can actually do.

    Article coming soon
  • KnowledgePIPE·04

    Inference infrastructure

    Model servers and APIs are services like any other: exposed, often weakly isolated, rarely rate-limited.

    Article coming soon
  • KnowledgePIPE·05

    Secrets & keys

    Model provider API keys and credentials easily end up in prompts, logs and pipeline variables.

    Article coming soon

Source of the LLM categories: OWASP Top 10 for LLM Applications 2026