CONTINUUM AI has launched a website containing AI agent incidents, distinguishing between actual damages and research cases.
Key points
•138 of the 354 incidents confirmed actual damage.
•The incident filter allows selection by 'actual damage status,' 'confirmation of AI involvement,' and 'record type.'
•51 incidents occurred in September 2026 alone.
CONTINUUM AI unveiled an AI agent incident archive website through OrcaRouter. The website includes 354 incidents and manages them by detailing whether actual damage occurred, the source, and the type of record.
The archive is categorized into three fields: 'real_harm,' 'ai_involvement,' and 'kind.' Of the total, 138 cases confirm actual damage, while the remainder are related to vulnerability disclosures, research, reports, and policy trends.
Additionally, the website provides filters that visitors can operate directly. Users can easily find desired information by selecting criteria such as 'actual damage status,' 'confirmation of AI involvement,' and 'record type.'
Analysis is also possible by attack type. The most frequent attack type is Credential Exploitation (CRED) with 55 cases, and 40 cases confirmed actual damage. Furthermore, 51 incidents occurred in September 2026 alone.
There are 2 records related to Korea, both covering AI security incidents in 2025 and 2026 under the policy category. The archive uses a regional field to indicate where the actual damage occurred, adopting a policy of presenting honest numbers rather than filling out a national page for Korean companies.
