LayersLayer 11 of 15
Safety and trust
Safety and trust covers the tools that check, test and control AI systems: evaluations, guardrails, security, red-teaming, monitoring, and the platforms that handle governance and compliance.
The categories
16 categories and 455 products in Safety and trust.
CategoryWhat it coversProducts
AI agent management platformsAI agent management platforms covers control planes that inventory, govern, secure, cost-manage and route an organisation's AI agents, whoever built them; separate from agent-building suites.20
AI governance & compliance platformsAI governance & compliance platforms covers systems of record for AI inventories, risk assessments, and regulatory compliance (EU AI Act, ISO 42001, NIST RMF).30
AI incident & risk databasesAI incident & risk databases covers public registries of AI failures, vulnerabilities, and harms used for threat intel and audits.14
AI-generated content & deepfake detectionAI-generated content & deepfake detection covers detection of AI-generated text, images, video and audio for moderation, education and fraud.28
Bias, fairness & explainabilityBias, fairness & explainability covers libraries and platforms to measure and mitigate bias and explain model decisions.34
Confidential AI & private inferenceConfidential AI & private inference covers platforms that run model inference, fine-tuning or agents inside hardware trusted execution environments (CPU/GPU confidential computing) with remote attestation, so prompts, data and weights stay encrypted even from the operator.32
Content provenanceContent provenance covers watermarking and content credentials that show where a piece of media came from.21
EvalsEvals covers tools and platforms that score model, RAG and agent outputs (offline suites, LLM-as-judge, online evals).37
GuardrailsGuardrails covers filters and policy checks on what goes into a model and what comes out.24
Interpretability & model internals toolingInterpretability & model internals tooling covers libraries and platforms for inspecting and steering LLM internals: activation hooks, sparse autoencoders, circuit analysis, feature steering.38
LLM & agent observabilityLLM & agent observability covers tracing, cost, latency, tool-call and usage analytics for LLM apps and agents.35
Moderation & guard classifiersModeration & guard classifiers covers models and APIs that classify text, image, audio or video for harm, jailbreaks, prompt injection and policy violations.33
Red-teamingRed-teaming covers tools and services that attack an AI system on purpose to find its weaknesses.30
Voice & conversational agent simulation testingVoice & conversational agent simulation testing covers platforms that simulate callers/users at scale, replay production calls and score audio-level behaviour (latency, interruptions, compliance) of voice and chat agents.14
Questions
Safety and trust, answered.
What is the Safety and trust layer of AI infrastructure?Safety and trust covers the tools that check, test and control AI systems: evaluations, guardrails, security, red-teaming, monitoring, and the platforms that handle governance and compliance.
What does the Safety and trust layer include?It has 16 categories: AI agent management platforms, AI governance & compliance platforms, AI incident & risk databases, AI security, AI-generated content & deepfake detection, Bias, fairness & explainability, Confidential AI & private inference, Content provenance, Evals, Guardrails, Interpretability & model internals tooling, LLM & agent observability, Moderation & guard classifiers, Prompt management, Red-teaming and Voice & conversational agent simulation testing.
How many Safety and trust products does Unfragile track?455 products across 16 categories, each category with its own page.
What works here
The three products on this layer.
SearchLive
RecommendNot live yet.
RouteNot live yet.