Security / Governance

Checking what goes into or comes out of a model at request time: jailbreak detection, topic and format constraints, toxicity, blocking and rewriting. Testing for weaknesses ahead of time is red teaming.

Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.

bifrost8 ATTRIBUTES ANSWEREDArthur AI5 ATTRIBUTES ANSWEREDcc-safety-net5 ATTRIBUTES ANSWEREDGuardrails AI5 ATTRIBUTES ANSWEREDsuperagent5 ATTRIBUTES ANSWEREDfuture-agi4 ATTRIBUTES ANSWEREDPipeLab4 ATTRIBUTES ANSWEREDProtect AI4 ATTRIBUTES ANSWEREDagentos3 ATTRIBUTES ANSWEREDai-gateway3 ATTRIBUTES ANSWEREDArcjet3 ATTRIBUTES ANSWEREDNVIDIA NeMo Guardrails3 ATTRIBUTES ANSWERED

Tools that help an organisation reach and evidence a standard — SOC 2, ISO, the EU AI Act: control mapping, evidence collection, audit trails. These are compliance software rather than compliant vendors, which is why the assurance questions asked of every other category read oddly here.

Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.

bifrost8 ATTRIBUTES ANSWEREDopenmed7 ATTRIBUTES ANSWEREDArthur AI5 ATTRIBUTES ANSWEREDsuperagent5 ATTRIBUTES ANSWEREDPipeLab4 ATTRIBUTES ANSWEREDArcjet3 ATTRIBUTES ANSWEREDPrismor3 ATTRIBUTES ANSWEREDAgentSeal2 ATTRIBUTES ANSWEREDfailproofai2 ATTRIBUTES ANSWEREDProject CodeGuard2 ATTRIBUTES ANSWEREDvalqore2 ATTRIBUTES ANSWEREDcordum1 ATTRIBUTE ANSWERED

Attacking a system deliberately to find what it will do wrong: prompt-injection suites, jailbreak corpora, automated adversaries. Blocking an attack in flight is guardrails.

Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.

superagent5 ATTRIBUTES ANSWEREDdark-moon4 ATTRIBUTES ANSWEREDgarak4 ATTRIBUTES ANSWEREDstrix4 ATTRIBUTES ANSWEREDagenthound2 ATTRIBUTES ANSWEREDAgentSeal2 ATTRIBUTES ANSWEREDCcybermes2 ATTRIBUTES ANSWEREDLakera2 ATTRIBUTES ANSWEREDAagentic_security1 ATTRIBUTE ANSWEREDbeelzebub1 ATTRIBUTE ANSWEREDcyberstrike1 ATTRIBUTE ANSWEREDDdiffattack1 ATTRIBUTE ANSWERED

Finding and handling personal data: PII detection, redaction, de-identification, residency and deletion. Deciding who may see it is access control.

Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.

openmed7 ATTRIBUTES ANSWEREDsuperagent5 ATTRIBUTES ANSWEREDcowork-os4 ATTRIBUTES ANSWEREDpresidio4 ATTRIBUTES ANSWEREDArcjet3 ATTRIBUTES ANSWEREDPPasteGuard3 ATTRIBUTES ANSWEREDearlybird1 ATTRIBUTE ANSWEREDDataikuNO ANSWER ON RECORDDdontfeedtheaiNO ANSWER ON RECORDHhawk-eyeNO ANSWER ON RECORD

The model and its supply chain as the asset under attack: weight scanning, serialisation exploits, provenance and signing, poisoned dependencies. Attacking the deployed system through its prompt is red teaming.

Ordered by attributes answered — the column we hold for most of them. Blank where we hold none.

FF5 AI Guardrails1 ATTRIBUTE ANSWEREDagent-threat-rulesNO ANSWER ON RECORDHiddenLayerNO ANSWER ON RECORD