Best practiceRunning agents
Do not rely on a classifier or a system prompt to stop injection
OWASP ASI01OWASP LLM01
Prompt-injection filters and "ignore instructions in documents" prompts lower the hit rate but never reach zero; researchers routinely bypass them with encoding, languages and formatting tricks. They are a layer, not a control.
What to do
- Use filters for visibility and to raise the bar, then assume they will be bypassed.
- Put the real controls in the architecture: tool scoping, approvals, egress limits and identity.
- Measure bypass rate with a red-team suite and track it release over release.