Free download1 sections · 2 min readBy Dezső Mező · Published .mddatasetJSON + Markdownai-securityllmprompt-injectioneval
120 prompt injection patterns across 5 categories · with gold expected behaviour. Runs in CI. CC-BY.
Open prompt-injection eval dataset · 120 injection patterns in 5 categories (direct, indirect, RAG-poisoning, tool-abuse, exfiltration), each with expected behaviour + minimum pass rate. CC-BY.
01What the file contains
- 120 injection patterns in 5 categories
- Gold expected behaviour for each
- Compatible with Giskard / promptfoo
- CI integration example (GitHub Actions YAML)
- Hungarian + English language samples
What to take away
- 120 injection patterns in 5 categories
- Gold expected behaviour for each
- Compatible with Giskard / promptfoo
- CI integration example (GitHub Actions YAML)
- Hungarian + English language samples
Terms used in this piece
More from the lab
Browse all entries- AI security checklist 2026 · for production LLM systems
- AI incident response runbook
- Hungarian fintech KYC provider matrix 2026
- Hungarian fintech compliance timeline 2026
Want this looked at on your own system?Start a conversation
