Compliance & audit
OWASP LLM Top 10 (2025) — mapping
Each of the OWASP LLM Top 10 2025 risks mapped to the AgentKeyRoute capability that mitigates it. Use this as a drop-in answer for RFPs and architecture reviews.
| Risk | Name | What it is | How AgentKeyRoute mitigates it |
|---|---|---|---|
| LLM01 | Prompt Injection | Hidden instructions in user input or retrieved content hijack the model. | Prompt-injection guardrail screens request and response; untrusted tool results are marked and re-screened; policy can require a safety-model pass before high-risk tools run. |
| LLM02 | Sensitive Information Disclosure | The model leaks PII, PHI, PCI, secrets or confidential business data. | Detection + redaction before data leaves; profile picks allow / redact / block per class; health, card and secret data never cached by default; response screening on the way back. |
| LLM03 | Supply Chain | Compromised models, datasets, extensions or plugins. | Signed Guard releases (sigstore), published SBOM, provider connections pinned per key; model identity verified so served-model mismatches fail closed. |
| LLM04 | Data and Model Poisoning | Attackers influence training or fine-tuning data. | Out of scope for an inference gateway, but the audit log preserves exact request/response pairs for later forensics and regression comparisons. |
| LLM05 | Improper Output Handling | Downstream systems trust model output as if it were safe. | Response guardrails can block, redact or mark output; the gateway labels the X-AIGov-Decision and X-AIGov-Model headers so downstream code can gate on them. |
| LLM06 | Excessive Agency | Agents call tools or take actions beyond what the user intended. | Workloads declare their autonomy (assistive vs autonomous); high-risk autonomous flows can be held for human approval; MCP tool-call policy limits which tools an agent may call. |
| LLM07 | System Prompt Leakage | An attacker extracts the system prompt and uses it to craft bypasses. | Content capture setting controls whether prompts are stored; system prompts can be injected server-side by the gateway rather than ever sent client-side. |
| LLM08 | Vector and Embedding Weaknesses | Retrieval returns confidential context to the wrong user. | Retrieval-side isolation is the app's job; the gateway can enforce that retrieval endpoints are only called by keys whose profile authorises that data class. |
| LLM09 | Misinformation | The model produces confidently wrong output that influences decisions. | Safety-model screen; response guardrails; workload tier marks high-risk uses for human oversight; profile can require citations. |
| LLM10 | Unbounded Consumption | A caller or looped agent consumes tokens, money or compute without limit. | Rate limits, token caps, workload budgets with hard enforcement, per-key spend ceilings, circuit breakers and queue cooldowns. |
How to configure a profile for OWASP coverage
- Pick a framework profile close to your use case (for example Internal assistant or Customer support assistant) and open its sensitive-data settings.
- Turn Prompt-injection sensitivity to Standard (or Strict for public assistants) under Guardrails (opens in a new tab).
- Enable a Safety model — the Automatic option picks one on your own providers.
- Under Workloads (opens in a new tab), set the correct autonomy and tier for each registered use.
- Set Monthly budget on each workload to defend LLM10 (Unbounded Consumption).
Verify coverage in the Request Inspector
Open any request in Requests (opens in a new tab) and read the decision trace: it lists exactly which OWASP class a block or redaction maps to.