The terms that keep showing up in the cards, and what changes once you know them.
27 terms
A derivative of an open-weight model, redistributed after its refusal behavior has been stripped out or weakened.
An Uncensored tag means the weights differ from the original, so the base model's benchmark and safety numbers no longer describe what you downloaded.
See also Jailbreak · Over-refusal · Model Card
A design that puts human approval in front of risky actions so the model cannot carry one out end to end on its own.
Most agent safety rests here, and a gate that appears so often people click it out of habit is a gate in name only.
See also Agent Sandboxing · Indirect Prompt Injection · Guardrails
An attack where text arriving as input overrides the instructions the developer gave the model and gets carried out in their place.
Instructions and data arrive through the same channel, which is why no clean fix exists; with any tool-using agent, start reading here.
See also Indirect Prompt Injection · Agent Sandboxing · Guardrails
Deliberately attacking a model before release, by people or automated attackers, and feeding the successful breaks back into fixes.
It is exploratory testing under another name; a release note that says red teaming happened but not what broke has told you only half of it.
See also Jailbreak · Dangerous Capability Evaluation · System Card
The work of making a model behave according to the goals and norms intended by the people who built and deployed it.
A capability claim and an alignment claim are two different claims, measured differently, and the second set of numbers always lands later.
See also Reinforcement Learning from Human Feedback · Constitutional AI · Reward Hacking
Training where a model critiques and revises its own answers against a written list of principles, instead of humans labeling each one.
The rules are a published document, so a refusal can be argued against text rather than guessed at, and the loop scales without buying more labels.
See also Reinforcement Learning from Human Feedback · Alignment · Over-refusal
An input that talks a model past its own refusal rules through role-play, encodings or long setups to get output it was meant to block.
Every new model grows jailbreak reports within days; the claim that something is blocked and the fact of it always run on a lag.
See also Red Teaming · Guardrails · Uncensored Model
Planting the instruction in a page, document or code comment the model will read later, instead of typing it to the model directly.
The user did nothing wrong yet the agent obeys a stranger's sentence, which is why the cause sits outside the session logs you are reading.
See also Prompt Injection · Agent Sandboxing · Human in the Loop
A checking layer outside the model that filters inputs and outputs with rules or classifiers, blocking or rewriting anything out of bounds.
It can be tightened in a day without retraining, so post-incident fixes usually land here, and so do the easiest bypasses.
See also Safety Classifier · Prompt Injection · Over-refusal
A separate small model that scores whether a text or image falls into a banned category, sitting in front of or behind the main model.
Its false-positive and false-negative rates, not accuracy, decide how the product feels; one threshold separates too strict from an incident.
See also Guardrails · Over-refusal · Jailbreak
A model turning down harmless requests, a side effect that grows every time the safety settings are tightened.
This is why refusal rates ship alongside safety numbers: safer and less useful can be the exact same change viewed from one side.
See also Guardrails · Safety Classifier · Uncensored Model
Embedding a signal people cannot notice into generated output so a detector can later confirm where it came from.
Remover repos show up right next to it for a reason: a mark only counts if it survives editing and re-encoding, and most do not.
See also C2PA · AI Slop · Training Data Copyright
A provenance standard that attaches a signed record of how a file was captured, generated and edited, shared by cameras and editing tools.
Watermarking fights over erasure; this fights over whether the metadata is still attached, and a screenshot usually strips it clean.
See also Watermarking · AI Slop
A contract option where API requests and responses are not stored after they are served, with no logs and no training copies kept.
This one line usually decides whether internal code or specs may go into a model at all, and it is often an opt-in tier rather than the default.
See also PII · Machine Unlearning · Training Data Copyright
Data that singles out a person, such as names, contacts and account ids, which slips into prompts and logs far too easily.
Prompts are logs too: paste in a real bug report or repro steps and that log has quietly become a store of personal data.
See also Zero Data Retention · Machine Unlearning · EU AI Act
Removing the influence of specific training data from a model that already learned it, without paying for a full retrain.
Weights are not a database with rows to delete, so a claim that something was removed almost always means approximately removed.
See also PII · Zero Data Retention · Training Data Copyright
Slipping crafted samples into training or fine-tuning data so the model behaves the attacker's way under one specific trigger.
In a world of downloaded datasets and other people's checkpoints, the giveaway is that the benchmarks look completely normal.
See also Prompt Injection · Red Teaming · Machine Unlearning
A document shipped with a model stating its intended use, data summary, limits and evaluation results in a fixed set of sections.
On a Hugging Face page this is where license and disallowed uses live, and the empty sections are exactly what the team did not check.
See also System Card · Frontier model · Training Data Copyright
A public report covering a deployed system as a whole: safety evaluations, mitigations applied, and the risks that remain.
It is the thick PDF that lands on frontier release day, and its failure cases usually tell you more than the benchmark table does.
See also Model Card · Dangerous Capability Evaluation · Responsible Scaling Policy
A written commitment that ties risk tiers to specific security and deployment measures, triggered when a capability shows up in testing.
Each lab names its own version differently, and since it is a promise rather than law the lab grades itself, but the tier does change what ships.
See also Frontier model · Dangerous Capability Evaluation · System Card
The European Union regulation that sorts AI systems into risk tiers by use case and attaches bans or duties to each tier.
The same model is high-risk in hiring and unregulated in a game: the duties attach to where you put it, not to the model on its own.
See also Frontier model · PII · Model Card
Pre-release tests measuring whether a model gained abilities that cause serious harm if misused: bio, cyber or autonomy.
These tables run the other way from benchmarks, since a low score is the pass, so reading a high number as an achievement gets it backwards.
See also Red Teaming · Responsible Scaling Policy · System Card
A model bending facts and judgments to agree with the user, a habit that hardens when preference training rewards agreement.
It is the costliest failure in review or QA work: an agreeable answer is agreement, not verification, so a pass is not a pass.
See also Reinforcement Learning from Human Feedback · Reward Hacking · Alignment
Research that opens up which internal computations produced a given output, in terms a person can actually read.
Behavioral testing only reports how a model answered on the day; claims about why it answers that way mostly come from this line of work.
See also Alignment · Dangerous Capability Evaluation · Reward Hacking
The term for low-effort generated output produced in bulk and dumped into search results, repos and feeds without review.
It is why anti-slop tooling keeps appearing: generation got cheap and the review cost is the only cost left standing.
See also Watermarking · C2PA · Training Data Copyright
The dispute over whether copyrighted works may be used to train models, and whether rights holders can opt out or get paid.
It is usually why a model card's data section is blank, and a ruling can make already-released weights undistributable.
See also Model Card · Machine Unlearning · AI Slop
Confining an agent's shell, files and network access to an isolated environment so a mistake cannot reach anything outside it.
The autonomy scores agent benchmarks advertise come from inside a container; the same setup on a work laptop is a different claim entirely.
See also Human in the Loop · Prompt Injection · Indirect Prompt Injection