# OpenA2A Research

> Security research on AI agent infrastructure. Exposure sweeps,
> behavioral threat reports, coordinated disclosures, and live indices.
> All findings classified by Threat Matrix technique. Methodology and
> raw queries published for every count.

This file is a discovery manifest for AI agents and autonomous
crawlers. For humans, start at https://research.opena2a.org. For
machines, the structured component index follows.

Maintained by: info@opena2a.org
Repositories: https://github.com/opena2a-org
License: research findings published under Apache-2.0; OASB benchmark
data and DVAA scenarios open source.

## Tracks

- Exposure Sweeps (monthly): infrastructure census via Shodan
  product-based queries. https://research.opena2a.org/reports/exposure-sweeps
- Behavioral Threat Reports (monthly): honeypot fleet observations
  classified by Threat Matrix technique.
  https://research.opena2a.org/reports/behavioral
- Coordinated Disclosures (event-driven): CVE pipeline publications
  following the 90-day disclosure protocol.
  https://research.opena2a.org/reports/disclosures
- First Observed In The Wild (event-driven): same-week posts when a
  pre-registered technique signature first fires.
  https://research.opena2a.org/reports/foitw
- Annual report: full-year consolidation across all tracks.
  https://research.opena2a.org/reports/state-of-ai-agent-security

## Live indices

- Exposure Prevalence: rolling counts of indexed AI services by
  category. https://research.opena2a.org/indices/exposure-prevalence
- Attack Prevalence: rolling counts of observed attack techniques
  across the honeypot fleet.
  https://research.opena2a.org/indices/attack-prevalence
- Sophistication: distribution of attacker skill markers, published
  distribution-only until a supervised classifier reaches the
  measurement threshold.
  https://research.opena2a.org/indices/sophistication

## Featured publications

- OASB Scanner Benchmark, April 2026: first published precision /
  recall / F1 for AI agent skill scanners. 89.2% F1 on 4,245 labeled
  samples across 9 attack categories. Comparison with 9 industry
  scanners.
  https://research.opena2a.org/reports/oasb-scanner-benchmark-april-2026
- Internet-Wide AI Exposure Sweep, April 2026: 321,929 AI-adjacent
  services indexed by Shodan as of April 10, 2026. First sweep using
  product-based queries.
  https://research.opena2a.org/reports/april-2026-exposure-sweep
- Internet-Wide AI Exposure Sweep, March 2026.
  https://research.opena2a.org/reports/march-2026-exposure-sweep
- The State of AI Agent Security: annual baseline.
  https://research.opena2a.org/reports/state-of-ai-agent-security

## Methodology

- Methodology index: https://research.opena2a.org/methodology
- Behavioral Sweep methodology:
  https://research.opena2a.org/methodology/behavioral-sweep
- First Observed In The Wild methodology:
  https://research.opena2a.org/methodology/foitw

## ARIA: Autonomous Research Infrastructure

Seven specialized agents producing the research above.
- About ARIA: https://research.opena2a.org/about
- ARIA architecture: https://research.opena2a.org/research/aria

## Methodology principles

- One classification, one resolver. Every observation maps to a
  Threat Matrix technique. https://threats.opena2a.org
- Reproducibility is the standard. Every count is reproducible from
  the public OpenA2A Registry telemetry export bundle within five
  percent within two hours.
- Distribution before mean. Subjective scoring ships
  distribution-only until a supervised classifier reaches the
  measurement threshold.
- Empty beats fabricated. An omitted cell beats an estimated number
  presented as measured.
- One number, one home. Numeric facts live on their spoke dashboard.
  A CI lint enforces this.

## Related properties

- Main site: https://opena2a.org
- HackMyAgent (CLI scanner): https://github.com/opena2a-org/hackmyagent
- DVAA (Damn Vulnerable AI Agent):
  https://github.com/opena2a-org/damn-vulnerable-ai-agent
- OASB (Open Agent Security Baseline):
  https://github.com/opena2a-org/oasb
- Threat Matrix: https://threats.opena2a.org
- AIM (Agent Identity Management): https://opena2a.org/aim

## Citation

If citing OpenA2A research in academic or industry publications,
prefer the report's canonical URL plus its publication date. Each
report includes a methodology section and the exact queries used.