Security from first principles

Take control of the agentic systems acting on your behalf

Lycid’s defense layers sit between your software and your AI to enforce security, performance and interpretability in high-stakes and regulated industries

Agent capability without blind trust

Injected content can enter an agent’s context. By architecture, it cannot reach tool execution or influence a decision unobserved.

Fast, real-time, reactive protection

Our latency measures in microseconds. Defences can adapt to the environment and previous interactions.

Direct insights on silent attack vectors

Per-decision observability: which sources influenced which action, with what trust class and weight — measured, not assumed.

Trustworthy AI by design

Capability-based policies are enforced before any action fires. Boundaries hold by structure, not by model judgement.

Uncompromised security

Probabilistic filters cannot guarantee protection. Lycid enforces security through a provenance graph — the record of which sources fed which action.

Every tool call, data dependency, and information flow is validated against a formal policy before execution.

Safety becomes a structural property of the system, not a best-effort heuristic.

  • Prompt injection contained by architecture, not filtered by heuristics
  • Deterministic policy decisions — no probabilistic false-positive trade-off
  • Security checks run in ~1ms
  • Fully deterministic
  • Auditable

Safe and interpretable decisions

Agent decisions are opaque by default. Lycid makes reasoning visible and structured.

Workflows become an explicit provenance graph: each node a tool call, transformation, or decision with tracked sources.

Inspect, audit, and constrain how an agent reasons before it acts with a verifiable formal structure.

  • Your agent can explain every decision it makes
  • Measure how much a decision moves under perturbations that should not move it
  • Adaptive reasoning patterns under varying levels of uncertainty
  • Route high-stakes decisions that draw on untrusted content to human review

Privacy and audit evidence

The AI is never touching your data, you can use a frontier model but keep everything local

Every value carries capability tags, its sources and who may read it.

Policies like "customer data may not leave the org"*" are enforced as graph properties, before a tool call fires.

  • Keep the AI and your data separate
  • Per-value source and reader attribution
  • Deterministic data-residency and exfiltration policies
  • Audit-ready evidence for the EU AI Act, GDPR, DORA and NIS2

Multi-agent systems

Every inter-agent message and MCP tool output enters the graph as an untrusted source.

Trust tags travel with messages, so classifications survive agent handoffs.

Policy gates fire on flows across the whole topology, not on string matching.

  • Agents can exchange graphs and permissions
  • The security boundaries naturally scale to multi-agent systems
  • Tool-description and MCP-server vetting & red-teaming

Small models

The AI thinks in logical languages that are concise and enforceable at generation, cutting both token budget and execution time.

The model that reads untrusted content can be small by design.

Plans are written once and executed deterministically. Security properties hold regardless of model size.

  • Keep token budget and execution time low
  • The logical structure of reasoning sets small models on par with larger ones
  • Lower cost and latency: plan once, execute deterministically
  • Structural security properties hold regardless of model size

Working with Lycid

We are onboarding a small number of design partners in regulated EU sectors for audits and security assessments of production or pre-production agentic systems.