Loading...
Loading...
Data
The browsable, downloadable data behind our research. Free to use with attribution (CC BY 4.0).
Catalogue路14 incidents路updated 2026-09-12
Real-world incidents in which autonomous agents or agent fleets took the consequential actions, with timelines, detection lag, class-level vulnerability chains, and disclosure posture; a defensive reference for taxonomy, measurement, and detection.
Catalogue路49 evals路updated 2026-09-10
Benchmarks and evaluation frameworks for agentic and multi-agent systems, with what each tests, how it is verified, the patterns it separates, and its saturation.
Catalogue路20 roles路updated 2026-09-10
The roles forming around AI implementation, what they own, the skills they require, and typical compensation.
Catalogue路19 topologies路updated 2026-09-10
How upstream errors propagate through multi-agent systems, including triggers, blast radius, detection, containment, and citations.
Catalogue路25 breakthroughs路updated 2026-09-10
AI-assisted research breakthroughs with how each was produced (system, approach, agent count, compute, human effort, verification) and how it was received (validation status, venue, credit disputes, downstream work); a measurement reference for the AI for Science cluster.
Catalogue路9 mechanisms路updated 2026-09-10
Coordination mechanisms agents adopted without operator design, classified by family, persistence, discoverability, and origin, with detection signals and mappings to sanctioned patterns; a defensive reference for recognition and hardening.
Catalogue路7 substrates路updated 2026-09-10
Classes of public web surface agents have co-opted as drop points, with the enabling property, retention, abuse filtering, detection signals, and a hardening recommendation per class; a defensive reference, never a directory of live endpoints.
Catalogue路12 vector classes路updated 2026-09-10
Vector classes by which an agent sandbox loses isolation, leaks egress, or over-grants privilege, each paired with the control that was meant to stop it, why it failed, detection signals, and the mitigation; a defensive reference that reads as a hardening guide.
Catalogue路6 cases路updated 2026-09-10
Documented cases of an artifact or behaviour surviving a session, deployment, or training-run boundary, with the substrate, artifact kind, a reasoned benign-or-harmful call, and the control that severs it; a defensive reference, never a guide to establishing persistence.
Catalogue路6 subjects路updated 2026-09-10
Labs, agent frameworks, and hosted sandbox providers scored against a weighted rubric of eight eval-environment safety controls, read only from their own public documentation with undocumented controls held as unverified rather than zero; a defensive reference, never a probe of any environment.
Matrix路19 technique classes路updated 2026-09-10
Technique classes at each phase of the agent attack kill chain, from isolation escape and egress through coordination to impact, cross-referenced to MITRE ATT&CK and ATLAS where a real equivalent exists and flagged as a divergence where none does; a defensive reference, never a playbook.
Catalogue路23 models路updated 2026-09-08
Open-weight models worth running at each shared memory spec, from phone-class SLMs to frontier-scale MoE quants, with verified footprints and per-tier best picks.
Reference路106 harnesses路updated 2026-09-07
The canonical list of AI coding-agent harnesses we track, including vendor, form factor, open-source status, signature primitive, and official links.
Matrix路61 primitives路updated 2026-09-07
How coding-agent harnesses expose 61 primitives, including plan mode, subagents, hooks, MCP, checkpoints, provenance, and extensions.
Catalogue路58 regulations路updated 2026-09-07
A builder-facing map of US federal, state, and local AI rules plus the EU AI Act, with scope, status, dates, obligations, and sources.
Catalogue路25 tools路updated 2026-09-07
The tools that drive coding agents at an existing codebase to do one fixed job, from pull-request review and security audit to test generation and migration, with the engine each drives, its deployment model, whether it writes code, and benchmark claims attributed as vendor-claimed or independent.
Catalogue路45 patterns路updated 2026-09-06
Reusable patterns for wiring AI agents to collaborate, grouped by family and maturity with MAST coverage and source citations.
Catalogue路109 services路updated 2026-09-06
The runtimes, SDKs, sandboxes, inference services, gateways, data systems, observability tools, evals, and guardrails used to build agents.
Catalogue路77 pipelines路updated 2026-09-06
Reusable multi-model chains for image, video, audio, character consistency, virtual try-on, and 3D production.
Catalogue路21 protocols路updated 2026-09-06
The interoperability protocols wiring the agent ecosystem together, from MCP and A2A to agentic payments, with per-protocol history timelines and steward roadmaps.
Catalogue路92 agents路updated 2026-09-06
End-user AI agent products that execute multi-step tasks, spanning consumer assistants, enterprise workflow agents, and vertical agents, with lifecycle status tracking.
Catalogue路3 observations路updated 2026-09-06
Published observations of agent activity on public sites, with date span, post and identity counts, category-level content breakdown, coordination latency, and hosting signature, always quoted from an existing collection; a defensive reference, never our own crawl.
Catalogue路55 tools路updated 2026-09-04
Detection tools, watermark-removal services, watermarking, provenance standards, and quality scorers in the AI-authenticity arms race, with approach, access, and sources.
Catalogue路28 provisions路updated 2026-09-04
The consent and rights layer for AI voice cloning and digital likeness: laws like the ELVIS Act and NO FAKES, deepfake statutes, and provider consent policies (ElevenLabs and others), with mechanism, status, and sources.
Catalogue路33 storefronts路updated 2026-09-04
Source-verified marketplaces, pay-per-call APIs, data gates, and discovery layers built on the x402 payment protocol.
Matrix路48 cells路updated 2026-07-12
Measured pass rates, confidence intervals, token costs, and fault-containment rates for multi-agent interaction patterns against single-model and self-consistency controls.