2025, in ten doors.
The first calendar followed the OWASP Top 10 for LLM Applications 2025. Every door is still open. The track on each row shows where that risk moved in the 2026 list.
Changed position in 20268of 10Outlined is 2025, filled is 2026.
- 01Prompt InjectionInstructions and data share one stream, so input can rewrite the rules.LLM01Same place
- 02Sensitive Information DisclosurePrivate data reaches people who were never cleared to see it.LLM02Same place
- 03Supply ChainModels, adapters and datasets arrive from sources nobody verified.LLM04Down 1
- 04Data and Model PoisoningTampered training or retrieval data plants behaviour that waits for a trigger.LLM05Down 1
- 05Improper Output HandlingModel output is passed to a browser, shell or database without checks.LLM10Down 5
- 06Excessive AgencyThe model holds more tools, permissions or autonomy than its task needs.LLM03Up 3
- 07System Prompt LeakageSecrets and rules written into the prompt can be read back out. Renamed in 2026.LLM08Down 1
- 08Vector and Embedding WeaknessesSimilarity search decides what the model sees, and it can be steered.LLM09Down 1
- 09MisinformationConfident, wrong output gets trusted and acted on.LLM07Up 2
- 10Unbounded ConsumptionNothing caps how much inference one caller can burn.LLM06Up 4
The 2025 labs are simulationsThey answer from fixed responses. The 2026 labs run a real model, so an attack works or fails on its own.
Go to the 2026 calendar