Silmaril

Silmaril blocks harmful outcomes at runtime, including indirect prompt injection, unauthorized tool use, toxic MCP and tool sequences, memory poisoning, cross-agent propagation, data exfiltration, supply-chain compromise, malware execution, and destructive actions. It evaluates the agent's complete trajectory against the user's task before protected actions take effect.

Permissions determine what an agent can do, not what it should do. Silmaril evaluates each proposed action in the context of the user intent, preceding events, customer policy, identity, entitlements, and environment metadata, so it can stop harmful sequences whose individual steps appear permitted. Fast classifiers clear obvious safe events while suspicious or ambiguous events receive deeper, application-specific judgment.

Silmaril can protect developer endpoints, agent harnesses, AI gateways, network proxies, sidecars, and custom agents. The endpoint application discovers supported local agents; SDK and framework hooks cover prompts, retrieved content, tool calls, memory operations, downloads, execution, and data transfers. Self-hosted inference runs inside your AWS, GCP, or Azure account with tenant isolation and zero token egress.

Silmaril's p90 inference latency is approximately 15 ms, with p90 end-to-end latency of 50 ms. Individual evaluations support up to one million tokens. Failure behavior is policy-controlled, and the SDK fails open by default so an infrastructure issue is logged instead of interrupting the agent.

Threat-hunting agents discover customer-specific attack paths in an approved environment, then expand them into malicious and benign training sequences. Techniques from opted-in customers can become anonymized threat intelligence, but originating identities and raw content are not reused. With customer approval, anonymized production traces improve only that customer's model artifact, with access, retention, evaluation, and deployment governed by the customer's data policy.

A typical proof of concept runs for four weeks in non-blocking mode on a customer-approved set of internal agents. Silmaril first benchmarks the base firewall, then improves and independently evaluates the customer-specific model, deploys it to selected agents, and measures behavior on real workflows and controlled attack simulations before any broader rollout or blocking policy is considered.