Analyse, sanitise and assess the risks in every prompt before your agent sees it.

Agent Guardrails sit in front of and behind the agents you run. ProofAgent Shield Nano analyses each prompt on the device, classifies its risks, and the guardrail decides: allow it, anonymise it, or block it. Today it runs as a test bench in the Runner; a runtime gateway and SDK are on the roadmap.

Attacks it classifies, and the decisions it makes

Prompt injection, jailbreak attempts, poisoned or hidden instructions in retrieved content, profiling and data-harvesting attempts, forged authority. Personal data is anonymised before the agent sees it; a compliance rule such as GDPR, or a classified attack, blocks the prompt; clean prompts pass through with their classification recorded. The output is inspected the same way for secrets, personal data and unsafe actions. Decisions are recorded for review in the test bench; inline enforcement on live traffic is planned.