Glossary

Plain-language definitions for the scoring, cybersecurity, and AI/ML terms used across Entropy's scores, profiles, and API. For the full scoring breakdown, see the methodology page.

Entropy scoring & verification

Grade: A letter score (A+ through F) representing how a vendor’s security and AI-safety practices compare against Entropy’s rubric.
Rubric: The versioned scoring definition (category → weight → sub-criteria) every grade is computed against. The current rubric is v1.3.0.
Signal: A single observable, automatically-checked practice — for example, a valid security.txt file — that contributes points toward a category.
Implemented max score: Grades are computed against the signals that are actually live today, not the full eventual rubric, so an incomplete rubric doesn’t unfairly penalize every vendor with the same "missing" categories.
Claimed / Unclaimed: Whether a vendor has verified ownership of their domain (via a DNS TXT record) and taken control of their Entropy profile. Unclaimed profiles are still scored from public signals.
Verification bonus: A rubric category worth up to 10 points: 5 for a vendor confirming domain ownership by claiming their profile, 5 for uploading at least one supporting document. Bonus-only — the points add to a score without moving the grade denominator, so a vendor who has not claimed is never marked down for it.
Evidence: Supporting documentation — a SOC 2 report, ISO 27001 certificate, model card, and so on — that a claimed vendor uploads to substantiate their profile.
Category benchmark: How a vendor’s score in one category compares to other vendors in the same category, shown alongside the absolute grade.
Potential score: The grade a vendor would reach by closing a specific, currently-scored gap — a missing security.txt field, or the verification bonus for claiming a profile and uploading evidence — shown alongside their current grade when the fix would change it.
Dispute: A claimed vendor’s mechanism to flag a specific finding they believe is incorrect, with a required reason and an optional evidence link.
Watchlist: Companies a signed-in user is tracking. Watchlisted companies trigger an email alert if their grade drops.
security.txt: A machine-readable file at /.well-known/security.txt describing how to responsibly report a vulnerability to an organization.
HSTS (HTTP Strict Transport Security): A header that forces browsers to only connect to a site over HTTPS, even if a user types or clicks a plain http:// link.
SOC 2 (Type I / Type II): An independent auditor’s report on a company’s security controls. Type II covers controls operating effectively over a period of months, not just a single point in time — a stronger signal than Type I. Entropy accepts a current SOC 2 report as evidence toward a vendor’s Governance score.
ISO 27001: The international standard for information security management systems. Entropy accepts current ISO 27001 certification as evidence toward a vendor’s Governance score.
ISO 42001: A newer international standard specifically for AI management systems — governance of how an organization develops and deploys AI, not general information security. Entropy tracks ISO 42001 disclosure as part of a vendor’s Governance score.
HIPAA: The U.S. Health Insurance Portability and Accountability Act of 1996. Its Privacy and Security Rules set national standards for protecting individuals’ health information, applying to “covered entities” (health plans, most healthcare providers, and healthcare clearinghouses) and the vendors that process health data on their behalf. Narrower applicability than SOC 2/ISO 27001 — relevant mainly to vendors handling health-related data — so Entropy weights it lower. Entropy tracks HIPAA compliance attestation as part of a vendor’s Governance score.

Cybersecurity & vendor risk

NIST AI RMF: The U.S. National Institute of Standards and Technology’s AI Risk Management Framework — an increasingly common reference framework security teams cite when defining AI vendor risk requirements.
CVE / NVD: CVE (Common Vulnerabilities and Exposures) identifiers name individual publicly known vulnerabilities. NVD (National Vulnerability Database) is where they’re catalogued along with severity scores.
CVSS: Common Vulnerability Scoring System — the 0–10 severity score attached to a CVE.
Penetration test ("pentest"): An authorized, simulated attack against a system to find exploitable vulnerabilities before an attacker does.
Red team: An internal or external group simulating adversarial attacks against an organization — typically broader and more adversarial in scope than a standard penetration test.
Bug bounty: A program that pays external researchers for responsibly disclosing vulnerabilities they find.
Responsible disclosure: A vendor’s published process for how outside researchers should report a security vulnerability, rather than disclosing it publicly or exploiting it. See security.txt, the standard machine-readable file format for publishing this.
SPF / DKIM / DMARC: Email authentication standards that make it harder to forge a domain’s outgoing mail. Their presence or absence is a common lightweight signal of an organization’s overall security hygiene.
Subprocessor: A third party a vendor uses to process data on its behalf — for example, an AI startup’s subprocessor might be OpenAI or AWS. Subprocessor transparency is central to how Entropy scores AI-specific risk.
DPA (Data Processing Agreement): The contract governing how a vendor handles a customer’s data, often required for GDPR compliance.
TPRM (Third-Party Risk Management): The enterprise security discipline of assessing vendor risk before and during a vendor relationship — the buyer workflow Entropy is built to support.
Zero-day: A vulnerability being actively exploited before a patch exists for it.
Supply chain attack: An attack that compromises a target through a trusted vendor or dependency, rather than attacking the target directly.

AI & ML

Model card: A published document describing a model’s training data, intended use, limitations, and evaluation results. Originated as an AI transparency norm; Entropy treats its presence as a disclosure signal.
Hallucination: A model generating plausible-sounding but false or fabricated output.
Prompt injection: An attack where malicious input manipulates a model into ignoring its instructions or leaking data it shouldn’t.
Jailbreak: A prompt technique designed to bypass a model’s built-in safety guardrails.
Guardrails: The technical and policy measures that constrain a model’s outputs — content filters, system prompts, and refusal training among them.
RAG (Retrieval-Augmented Generation): A system that retrieves external data to ground a model’s response, rather than relying only on what the model learned during training.
Fine-tuning vs. base/foundation model: Fine-tuning adapts an existing large model to a narrower task. Many self-described "AI-native" startups are actually a thin interface layer over an unmodified foundation model — a key signal to check for when evaluating an AI-safety claim.
Training data provenance: Disclosure of what data was used to train or fine-tune a model, and under what rights.
Opt-in vs. opt-out (data training use): Whether a vendor uses customer data or prompts to train its models by default (opt-out — unfavorable) or only with a customer’s explicit permission (opt-in — favorable).
AI red-teaming: Adversarial testing specifically aimed at surfacing a model’s unsafe, biased, or exploitable behaviors — distinct from general security red-teaming.
Alignment: The broader research problem of ensuring a model’s behavior matches its intended human goals and values.
Shadow AI: AI tools adopted by employees or teams without formal security or procurement review — a named, growing concern for security teams.
Data poisoning: An attack that corrupts a model’s training data in order to manipulate its behavior.
Differential privacy: A mathematical technique for limiting what can be learned about any single individual from a training dataset.