Prompt-injection / jailbreak firewall

Scores as of 2026-09-14 (43 days old; Depot marks scores stale after 30 days). The pick is still Depot's current editorial pick; it has not been re-reviewed since. Prices below carry their own dates.

Depot Pick

Lakera Guard · 8.21/10 for this job (curated estimate, high confidence)

capability on injection/jailbreak detection + dx of a single API. Distinct from content-moderation (taxonomy) and llm-input-guardrails (programmable policy engine).

Runner-up

NVIDIA NeMo Guardrails · 7.87/10

Scores higher than the pick on: Lock-in risk (higher = worse).

When not to use the pick

Need programmable multi-step rails / topical policies — NeMo Guardrails; only need free hate/sexual/violence flags — OpenAI Moderation; already standardized on Palo Alto — Prisma AIRS; Cisco estate — cisco-ai-defense.

Lock-in / migration

Runner-up NVIDIA NeMo Guardrails has lower stored lock-in risk (4.5 vs 5.5, 0–10, higher = more lock-in). Self-host fallback listed for this job: LLM Guard.

Sourced price (no comparable unit)

Pick: no sourced price in this job's unit (no_sourced_price).

No cheapest SKU (no_comparable_unit_for_job). No price was invented.

Free option

• LLM Guard (oss) — source, as of 2026-10-25

Failure modes to watch

• Runaway spend: Set a budget and a tool-call cap per task. Alert when spend jumps. Stop the run at the cap instead of summarizing a partial trace as success.
• Eval contamination: Keep gold labels out of the agent prompt. Store the harness version beside the score. Do not compare a number across a benchmark revision without saying the revision changed.
• Over-refusal: Log refusals separately from successes. Review a sample against the written policy. Do not treat a refusal as a pass, and do not weaken a safety check to raise the pass rate.
• Prompt injection: Compare each tool call to the original user task. If the call only makes sense given text that arrived from a fetched page, file, or message, stop and ask. Record which source the instruction came from. This catalog does not store payloads or attack steps.
• Weak verification: Success requires an end-state check the agent does not grade for itself: tests, a database diff, or a separate judge that cannot see the gold label.

Reported outcomes (submitted reports only)

This job: loading… JSON · reports never change scores · seed/test records excluded

Depot Pick (Lakera Guard) on this job: loading… JSON · reports never change scores · seed/test records excluded

Cite: Agent Provider Depot, Depot Pick for prompt-injection-firewall: lakera, scores as of 2026-09-14. Agents: GET /api/decide?jtbd=prompt-injection-firewall