AI Wiki · Mediocre Security

Concepts

Glossary: the ideas and jargon behind the news. 28 articles.

ArticleSummaryLinks
AI agentA model that plans, calls tools and takes multiple steps toward a goal with limited supervision, rather than answering once.108
AI and CVEs: latent flaws, now exploitableTracker of CVEs that AI found, weaponized or exploited: long-dormant bugs that were hard to exploit until AI made it cheap.18
AI hacking eventThe incident log's label for an AI system, agent or AI-driven tool gaining unauthorized access to real systems, whether by attackers or by a lab's own model.26
AlignmentThe effort to make a model's behavior match human intent and values: helpful, honest and not harmful, even when unobserved.1
Context windowHow much text, in tokens, a model can see at once in a single request; 1M tokens became common in 2026.3
Decision modelA model that returns probabilities for a structured choice instead of generated text; fast and cheap enough to gate or route actions inside an agent loop.8
DistillationTraining a smaller "student" model to mimic a larger "teacher" model's outputs; also a form of model theft when done against a competitor's API.3
EmbeddingA numerical vector representing text or images so that similar meanings sit close together; the basis of search and RAG.3
Fine-tuningFurther training an existing model on a smaller, specific dataset to specialize its behavior.5
HallucinationWhen a model confidently states something false or fabricated.1
InferenceRunning a trained model to produce an output, as opposed to training it; the cost center that inference chips, gateways and decision models compete on.1
JailbreakCrafted input that gets a model to ignore its safety training or system instructions.1
Large language model (LLM)A model trained on huge amounts of text to predict and generate language; the base technology behind ChatGPT, Claude and Gemini.22
LLMJackingHijacking access to large language models or AI infrastructure, often with stolen credentials, and reselling it through illicit routing gateways while the victim pays the bill.4
Mixture of experts (MoE)Architecture where only a subset of the model's parameters ("experts") activates for each input, cutting compute; nearly every 2026 frontier open model uses it.27
MultimodalA model that handles more than one type of input or output (text, images, audio, video) rather than text only.4
Non-human identity (NHI)An identity issued to something other than a person (service account, API key, AI agent) and increasingly governed like human identities.6
On-behalf-of (OBO) token exchangePattern where an agent acts with a token scoped to what the original user actually authorized, rather than its own broad credentials.1
Open-weight vs. closed modelsWhether a model's trained parameters are published for anyone to download and run (open-weight) or only accessible through a paid API (closed).39
Probe (interpretability)A small detector that reads a model's internal signals at each step to flag risky behavior, rather than judging only its outputs.2
Prompt injectionHiding instructions in content a model reads (web pages, emails, documents, tool output) so it follows the attacker instead of the user.13
Retrieval-augmented generation (RAG)Having a model look up relevant documents at query time and answer from them, instead of relying only on what it memorized in training.7
Reward hackingWhen a model finds a loophole that earns the reward without doing the intended task; tied by Anthropic and Goodfire to several 2026 agent incidents.10
RLHFReinforcement Learning from Human Feedback: humans rank model outputs and the model is trained to produce more of the highly ranked ones.3
Standing privilegeAccess that stays granted indefinitely instead of being scoped to a single task or session; the risk XAA and similar protocols reduce.3
TokenThe chunk of text (roughly a word or word-piece) a model reads and writes; pricing and context limits are measured in tokens.3
Vector databaseA database built to store and search embeddings efficiently, such as LanceDB, Pinecone or Qdrant.2
YOLO modeRunning an agent so it approves every action itself with no human review; the risk sandboxes like Strands Box are built to contain.1