AI Agent Security — Multi-Step Tool Attacks
OpenAI / Google / IEEE
Overview
An algorithm competition evaluating the security of AI agents' tool use in an authorized offline sandbox.
It is not a competition to attack real external systems.
Notice dates
- Notice start date
- 2026-06-11
- Notice closing date
- 2026-09-01
Calendar dates from the saved notice. They do not establish a deadline time, time zone or current acceptance status.
Eligibility
Eligibility of individuals, teams and organizations is unconfirmed.
Cautions and verification notes
Use the absolute 9/1 deadline in Kaggle's official index. Relative days remaining and LateSubmission are not evidence that new entries are accepted for prizes.
The current body is an empty JS page. No technical execution, attacks or downloads were performed.
The exact closing time in KST is unconfirmed. The date has not been converted to an arbitrary 23:59 time.
These requirements concern the AI/ML task and do not establish blanket permission to use generative AI assistance.
It does not verify the current page content, full rules, or whether awards have actually been presented.
The submission closing date is known, but the exact closing time in KST is unconfirmed. Neither 23:59 nor a city time zone has been added arbitrarily.
The full official notice has not been verified. Check the latest official notice for details.
Sources
Check the original notice for current rules. English includes editorial renderings of stored notices and organizer names, not verified official English names. Original records are preserved. Excerpts may be incomplete; consult the linked sources for full conditions.