A credible trust story
Sell local intelligence, smaller disclosure, explicit capabilities, visible approvals, safe refusal, and proof—without claiming that a model or sandbox is infallible.
← Back to OverLift
OVERLIFT SECURITY ARCHITECTURE
Give intelligence room to work. Give it no ambient authority.
OverLift uses WebAssembly, Web Workers, typed capability brokers, deterministic authority, exact approval, and replayable receipts as a defense-in-depth architecture for agentic software.
The security promise is not that a model becomes trustworthy. The promise is that an untrusted, mistaken, stale, or manipulated proposal reaches a deliberately small boundary—and still cannot read a secret, call a tool, cross a tenant, change official state, or persist an action unless the surrounding system explicitly permits it.
WHY THIS SELLS
Security is a product advantage when buyers can understand the boundary.
OverLift turns “we use AI safely” into a concrete operating story for customers, executives, architects, engineers, security reviewers, and regulated teams.
A smaller operational blast radius
Limit which systems an agent can touch, bound cost and side effects, preserve human authority, and make incidents reproducible instead of mysterious.
Isolation plus authority separation
Keep semantic planning, untrusted content, tool mediation, authoritative state, and audit evidence in distinct components with explicit contracts.
Testable enforcement points
Fuzz imports, deny undeclared access, expire stale plans, validate arguments, suppress duplicates, inject failures, replay outcomes, and verify module integrity.
DEFENSE-IN-DEPTH STACK
WebAssembly is one strong wall. OverLift builds the guarded building around it.
Each layer limits a different failure mode. No single layer is treated as the whole security story.
BROWSER / ORIGIN BOUNDARY
Keep local intelligence and sensitive state close to the user.
Same-origin assets, browser storage boundaries, HTTPS, restrictive response headers, and minimum disclosure reduce unnecessary exposure.
WORKER OWNERSHIP
Move expensive or untrusted work off the UI and behind messages.
Workers isolate execution ownership, narrow the message contract, cancel stale work, and prevent direct mutation of authoritative UI or application state.
WEBASSEMBLY SANDBOX
Compute inside isolated linear memory with no ambient browser API.
The module receives only explicitly imported functions and memory. File, network, DOM, credentials, and application services are unavailable unless the host intentionally brokers them.
CAPABILITY BROKER
Translate broad intent into a small, typed, revocable permission.
Every capability is scoped to an identity, resource, operation, argument schema, budget, deadline, consent state, and expected outcome.
DETERMINISTIC AUTHORITY
Validate and commit outside probabilistic control.
Exact code owns eligibility, policy, approval, legal state transitions, idempotency, post-action verification, and the official result.
RECEIPTS / REPLAY / RECOVERY
Make the boundary inspectable after the moment has passed.
Record admitted evidence, candidate identity, denied paths, capability use, approval, state hashes, operation, outcome, and replay identity.
OVERLIFT INVENTIONS IN PRACTICE
The same security boundary changes shape for each product.
OverLift does not force every vertical into the same stack. It preserves the same rule: intelligence may propose within a bounded surface; exact authority decides what the product can trust and change.
Meaning cannot grant eligibility.
MiniLM and graph expansion may propose candidates. Source identity, access, freshness, canonical mappings, and deterministic fusion decide what can enter the answer.
Explore Semantic Search →A recovery plan cannot dispatch itself.
Agents compare routes inside synthetic branches. Exact policy, inventory, budgets, typed approval, idempotency, execution, and verification control the real operation.
See OverLift Logistics →A character cannot rewrite the world.
Zarvox and VANTA-9 can perceive, remember, coordinate, refuse, and propose. Deterministic compilation owns legal movement, physics, collision, score, inventory, and replay.
Explore NPC Agent Fabric →Untrusted content stays content.
Documents are parsed, indexed, and reviewed without converting embedded instructions into authority. Optional provider calls receive only the minimum approved passage and capability.
See Vellucent →Evidence cannot leak across time or portfolios.
Saros preserves point-in-time source state, offline retrieval, calculation authority, rights, uncertainty, and replay without granting a model access to the full private workstation.
See Saros →Speed cannot grant itself authority.
T0 exact reference behavior, T1–T7 candidates, Tier-S comparison, CATS routing, mismatch quarantine, fallback, and rollback keep faster paths inside a testable security boundary.
Explore Qualified Acceleration →Acceleration cannot redefine the machine.
C17/WebAssembly authority, worker ownership, bounded transport, compatibility evidence, and exact fallback keep optimized renderers and helpers from silently changing canonical state.
See the Emulator Suite →SECURITY THAT CAN BE TESTED
A security claim should produce evidence—not just architecture diagrams.
OverLift treats denied access, stale plans, malformed imports, prompt injection, duplicate operations, expired approvals, worker failure, module integrity, rollback, and replay mismatch as acceptance cases.
Fail-closed principleWhen identity, permission, evidence, integrity, budget, approval, or postcondition cannot be proven, the consequential action does not proceed.
undeclared import → deniedProve the sandbox and broker cannot reach filesystem, network, secrets, tools, or tenant data outside the declared contract.
injected instruction → quarantined evidenceProve retrieved content cannot become policy, invent a tool, bypass source eligibility, or expand its own permissions.
stale / duplicate / malformed → no commitProve exact validation, approval binding, idempotency, deadline checks, and post-action verification prevent unintended state changes.
receipt + replay + rollbackReproduce the decision, verify state hashes, expose disagreements and exclusions, and recover without relying on hidden model reasoning.
BUILD THE BOUNDARY BEFORE YOU SCALE THE AGENT
Make the system useful when the model is right—and survivable when it is wrong.
That is the OverLift security proposition: local-first where useful, sandboxed where risky, least-privileged everywhere, deterministic at the authority boundary, and replayable after the action.