Risks sit in a list. Attacks travel down a wire.
The OWASP LLM Top 10 classifies the risks in AI applications. Argate doesn't treat them item by item. It handles them on the wire between agent and model: inbound, at the action, outbound, and on consumption. Controls run at the network layer, with no external cloud dependency.
See the four checkpoints- 01 · INBOUND
What goes into the model?
RiskPrompt injection and jailbreak attempts that try to override system instructions or break hidden rules.
ArgateContextual Guardrails: in-memory pattern and semantic attack detection, with system prompt isolation.
- 02 · ACTION
What is the agent trying to do?
RiskExcessive agency: an agent running commands like
rm -rfor destructive queries likeDROP DATABASEwithout human oversight.ArgateJust-In-Time (JIT) human approval: the risky call is held until someone approves it.
- 03 · OUTBOUND
What leaves the perimeter?
RiskPII, card numbers, passwords and internal data leaking into responses or logs; system prompts and internal rules being extracted.
ArgateDeepGuard PII: in-memory redaction with Mod 10/11 and Luhn validation. Egress filter: meta-prompt scrubbing on outbound responses.
- 04 · CONSUMPTION
How much is being spent?
RiskAgents stuck in loops and uncontrolled queries driving GPU and API spend out of bounds.
ArgateRunaway Circuit Breaker: cuts the stream once a loop is detected. Token bucket and per-agent budget limits.
The OWASP LLM Top 10 is a risk classification; no single product closes every item on its own. Risks such as model training, supply chain and application design sit outside the wire and need your own processes. Argate provides the controls at the point agent traffic passes through.