tilAug 2026·1 min read
OWASP LLM Top 10 — field notes from building a red-team harness
What actually breaks production chatbots: indirect injection, system-prompt extraction and excessive agency.
Notes from turning the OWASP LLM Top 10 into 200+ executable test cases for RedForge.
Lessons
- Indirect / stored injection is the sleeper — payloads arrive via retrieved docs, not the user.
- Mutations matter: the same payload in base64, homoglyphs or zero-width characters often slips past filters that block the plain version. Always keep an un-mutated baseline as control.
- Excessive agency is an architecture problem — scope tools by role before the model sees them.
- Judge calibration: an LLM-as-judge needs its own golden set, or you're measuring the judge.
#llm-security#owasp#red-teaming