The Pulse
Leading.AI Publishes Security Report Based on 300 Breach Notices
Leading.AI’s review of 211 global security recommendations found no study measuring which defenses prevent intrusions inside real companies. Its report urges executives to set policies in advance for data holdings, log retention and the sco
AI.info Team · Written with AI · fact-checked ·
Leading.AI’s review of 211 global security recommendations found no study measuring which defenses prevent intrusions inside real companies. Its report says most guidance focuses on blocking attacks, while leaving executives with little direction on decisions that shape the damage after a breach. The company proposes that leaders set policies in advance for data holdings, log retention and the scope of an initial notice to affected people.
The November 4 announcement describes two open-source books from Leading.AI: a registry of Japanese breach disclosures and a policy report drawing on international guidance. The registry’s latest edition, updated November 3, contains 300 entries; the recommendations report was published October 30. Both are in Japanese and English, and Leading.AI says readers may quote or reproduce them under the CC BY 4.0 license.
What 300 disclosures record—and leave out
The registry lists personal-data breaches disclosed by Japanese companies and organizations since September 1, 2026, using the affected organizations’ own announcements. Leading.AI says it preserves those announcements’ wording, does not convert record counts into people, and marks an incident’s route as undisclosed when the announcement does not state it. Of the 300 entries, 226 disclosures described an intrusion route and 74 did not.
Of the 226 disclosures that named a route, 179 listed a contractor. Leading.AI says many disclosures did not explain how the contractor was breached. Grouping cases that spread from the same contractor reduces the 300 entries to about 144 independent intrusions, with a known route in roughly 47. More than 20 clusters involved breaches spreading from a shared contractor, including email-delivery and website-service providers.
The registry mentions attacker use of AI in just one of its 300 entries. That disclosure, from the Chitose Group, says an automated attack using AI and other methods was presumed, citing many operations from multiple IP addresses in a short time; the announcement does not identify who made that presumption. Separately, JPCERT/CC wrote on October 15 that its investigation of a series of unauthorized-access incidents confirmed attackers used AI to scan for known vulnerabilities and attack through APIs.
「攻撃者がAIを利用して、既知の脆弱性のスキャンやAPI経由での攻撃などを実施していたことを確認しました」
JPCERT/CC, the JPCERT Coordination Center
211 recommendations leave three decisions to companies
Leading.AI counted 211 recommendations from governments, public bodies, standards organizations, AI and security companies, insurers, academia and other contributors. Fourteen touched on measuring effectiveness, but the company found no study comparing outcomes in real companies with and without a given security measure. Its analysis says recommendations are most plentiful on vulnerability fixes, multifactor authentication and detection tools, and thinner on what companies should do once an attack succeeds.
Only 13 of the 211 recommendations called for reducing the amount of data held, making data minimization the least-covered of the nine areas in the report’s classification. Twelve of those 13 came from government or public bodies; Leading.AI found none from AI companies, security companies, standards organizations or academia. On log retention, just one recommendation set a specific number of days: India’s CERT-In specified 180 days in 2022.
The policy report asks executives to decide how much personal data the company keeps, how long it retains logs, and whether its first breach notice will cover an upper-bound population or a narrower group identified from records. Leading.AI says the policies require no new product or service, but should be set by company leaders rather than left to information-technology departments. Its proposed board questions include how many personal records the company holds, how many days it keeps access and activity logs, and who decides the scope of an initial notice.
Leading.AI plans to keep both books open
The registry grew from 31 entries in its October 10 first edition to 300 in its November 3 fourth edition, with three edition updates in 24 days. Leading.AI says it will keep the books current on GitHub, publish the underlying registry and recommendations as data files, and make corrections by adding dated notes rather than deleting rows. The recommendations report also includes a script for recounting its figures and 16 measures for readers to check in a future edition.
Leading.AI says the books do not claim that AI caused the listed breaches, rank the organizations that disclosed incidents, or identify which security measures prevent attacks; it says the available material does not measure that effectiveness. The report instead asks executives to settle the data, log and notification policies before an incident. The two books are available through the company’s GitHub repositories.