- Capability group
- Artificial Intelligence
- Domain
- Critical Technology · Hard Power
- Identifier
- factbase.org/cap/T266WRYB20
- Reference
- Adversarial machine learning Wikipedia
Papers
32.2k
2020–2024
Paper growth
+750%
2020–24 on 2015–19
Patent families
2.3k
2020–2024
Filed abroad
37%
families in more than one office
In the framework
What the capability covers and what it leaves to others, and the indicators it will be measured by.
Bounds
Covers the technical safety, security and evaluation capacity; regulation and international governance belong to the governance domain, and cyber defence of AI systems is shared with the protective cybersecurity capability.
Lineage
Carried from v0.0 (formerly Adversarial AI) with the same identifier; broadened from attack-and-defence research to the full evaluation capacity, matching the institutionalisation of model assessment as a standing state function.
Indicators
- research publication share in AI safety and evaluationResearch
- standing national model evaluation institutesPending
- frontier model evaluations publishedPending
Research
Papers classified to the capability, by year and by the nation of their authors' institutions, 2020–2024. Read from the v0.0 research topic Adversarial AI.
Papers by year
The five leading nations, papers by year
Share of papers, 2020–2024
Of papers with a known author nation; 41% have none and are left out.
- 1China29.9%6,824 papers
- 2United States21.5%4,904 papers
- 3India4.8%1,100 papers
- 4United Kingdom4.2%959 papers
- 5Australia3.5%800 papers
- 6Canada3.1%698 papers
- 7Germany2.9%662 papers
- 8Singapore2.3%524 papers
- 9South Korea2.2%512 papers
- 10Japan2.2%508 papers
Leading institutions, 2020–2024
By fractional count: each paper divided among its authors' institutions.
- 1Tsinghua University171fractional papers
- 2National University of Defense Technology166fractional papers
- 3University of Electronic Science and Technology of China155fractional papers
- 4Purdue University West Lafayette151fractional papers
- 5Shanghai Jiao Tong University151fractional papers
- 6Nanyang Technological University150fractional papers
- 7Xidian University145fractional papers
- 8Beijing University of Posts and Telecommunications137fractional papers
- 9Wuhan University129fractional papers
- 10Zhejiang University125fractional papers
Patents
Patent families by the nation of origin of their applicants and their earliest priority year, from the CPC and IPC classes the capability is read from. Read as tools for the field: the classes count inventions used in it rather than the field itself. 2025 onwards is incomplete: families are published 18 months after priority.
Families by priority year
The five leading origins, families by year
846 applicants filed here in 2015–2024; the five largest held 14% of families (HHI 69 of 10,000: unconcentrated).
Share of families by origin, 2020–2024
- 1China44.9%1,020 families
- 2United States25.0%569 families
- 3Japan9.9%226 families
- 4Germany5.4%123 families
- 5South Korea5.4%122 families
- 6Canada1.1%24 families
- 7Sweden0.9%20 families
- 8France0.8%19 families
- 9Ireland0.8%18 families
- 10United Kingdom0.6%13 families
Applicants filing most internationally, 2015–2024
Families filed in more than one office, by applicant as the offices name them.
- 1FUJITSU LTD73international families
- 2NEC CORP46international families
- 3SIEMENS AG40international families
- 4BOSCH GMBH ROBERT37international families
- 5HUAWEI TECH CO LTD36international families
- 6IBM33international families
- 7HITACHI LTD29international families
- 8MICROSOFT TECHNOLOGY LICENSING LLC28international families
- 9ERICSSON TELEFON AB L M20international families
- 10SONY GROUP CORP14international families
Classification read
- G06N5/045 Explainable and interpretable AI · 10.8k documents
Shares inventions with
Capabilities whose classes the same families also fall in, by overlap (Jaccard).
- Machine Learning & Foundation Models1,011 families · 0.009
- Applied AI & Advanced Analytics985 families · 0.005
Companies
50 companies in this capability, across 5 regions. The leading 12:
- 1OpenAIFrontier developer that runs a Preparedness Framework, external red-teaming network and published system-card evaluations, and grants pre-deployment model access to the US CAISI and UK AISI.Research
- 2AnthropicFrontier lab built around alignment and interpretability research; it runs a Responsible Scaling Policy with capability-threshold evaluations and open-sourced evaluation tooling.Research
- 3Alphabet (Google DeepMind)Google DeepMind's Frontier Safety Framework, dangerous-capability evaluations and SynthID watermarking, plus Google's Secure AI Framework (SAIF), make it a top safety and evaluation actor.Research
- 4MicrosoftRuns a dedicated AI Red Team, the open-source PyRIT red-teaming toolkit and Azure AI Content Safety/Prompt Shields, and is a leading patent filer in this field.Software & services
- 5Meta PlatformsPublishes open safety tooling for open-weight models (Llama Guard, Prompt Guard, CyberSecEval in the Purple Llama project) under its Frontier AI Framework.Research
- 6Scale AIIts SEAL research lab runs leaderboards and red-teaming and evaluation services for frontier labs and US government (Meta took a 49% stake in 2025).Software & services
- 7NVIDIASupplies the open-source NeMo Guardrails, the Garak LLM vulnerability scanner and an AI red team, which are widely used to secure deployed models.Software & services
- 8IBMCreated the Adversarial Robustness Toolbox and sells watsonx.governance and Granite Guardian; it is among the top international patent applicants here.Software & services
- 9Palo Alto NetworksBought Protect AI in 2025 and folded its model scanning and LLM red-teaming into the Prisma AIRS AI-security platform.Software & services
- 10Cisco SystemsBought Robust Intelligence, a pioneer of AI model validation and algorithmic red-teaming, in 2024, and now sells it as Cisco AI Defense.Software & services
- 11Check Point Software TechnologiesBought Swiss LLM-security firm Lakera (Lakera Guard, Gandalf red-teaming dataset) in 2025, which anchors its AI security offering.Software & services
- 12FacultyRuns pre-release safety and red-teaming evaluations of frontier models for labs and the UK AI Security Institute; Accenture agreed to acquire it in January 2026.Software & services
