Block prompt override attempts in output.
This page is the SEO reference for the prompt_injection validator. Configure it on an agent in the dashboard or via the agents API. Full catalog: validators.
What it checks
Block prompt override attempts in output.
| Field | Value |
|---|---|
| Rule ID | prompt_injection |
| Pack | core |
| Category | Security |
| Default severity | error |
| Default on_fail | block |
Config notes
Attach the rule to an agent, set severity and on_fail, then start in shadow mode before enforce. Pack core must be allowed on your plan.
Pass / fail example
Pass: output satisfies the encoded contract for this rule (shape, pattern, or checksum as configured).
Fail: output violates the contract. ValGuard returns a validation failure with this rule ID when on_fail is block or reask.
Exact fixtures depend on your field paths and thresholds. Use the dashboard tester or validate LLM output.
on_fail behaviours
| Mode | Effect |
|---|---|
block | Request fails closed; next step should not run |
reask | Proxy may retry the model once within plan limits; buffers stream |
warn / log | Traffic continues; audit still records the evaluation |
| shadow agent | Would-block logged; response still delivered |
Prefer fail-closed for money and PHI paths.
Performance
Engine timing: 9.6 µs engine p50 (snapshot). HTTP path and model time dominate end-to-end latency. See methodology. Numbers are ValGuard's own measurements on published hardware snapshots.
Related
Next step
Quickstart. Attach prompt_injection in shadow on one agent, then enforce.