85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Applied AI lab

Safer, reliable agentsfor production.

Motiontic is an applied AI lab building safer, more reliable AI agents for enterprises: a standard harness SDK, security tooling for the agentic stack, and production agents built on both.

Agentic tools audited
85
Companies shipped with
11
Security incidents logged
221
First production code
Week 1

85 agentic tools audited for security incidents · full audit coming soon

Eleven companies across AI infrastructure, fintech, revenue software and retail tech.

NeevCloudFringFlo FinanceQuantum HeapsPulseGM MarketsDrevonNarrativemirrARDurbinQuinch Systems

Two problems,01in every agent build

Across agent builds for real businesses, the same two problems came back. Both are infrastructure problems, not model problems.

Problem 1

Every team builds its harness from scratch.

Each team writes its own loop, tool calls and retries, then spends months making that one harness reliable, starting from zero.

A harness should be a standard SDK: tested once, hardened over many deployments, and reused across very different agents.

Problem 2

Few agent SDKs are production-ready. Fewer are built for the web.

Outside Anthropic, OpenAI, LangGraph and a small number of others, few agent SDKs hold up in production.

Fewer still are designed for web software, which is where most enterprise agents have to run.

85 tools, audited02for security incidents

Agent SDKs, harness SDKs, MCP servers, skills, sandboxes and gateways, reviewed one by one. 68 of the 85 had at least one incident in 2025–26.

tools had at least one security incident in 2025–26
68 / 85
security incidents across the 85 tools
221
of those incidents happened in 2026
166
tools where GW-2 applies had a relevant CVE
51 / 70
GW-2

The most common risk in the audit. Treat frameworks and gateways as critical supply chain.

Full audit coming soon
In scope:Agent SDKsHarness SDKsMCP serversSkillsSandboxesGateways

What the lab03is building

Three tracks, each one a direct response to the field work and the audit. Reusable, tested and secure by default.

SDK

A standard harness

Loop, tools, memory, retries and guardrails. Written once, tested across client deployments, reused for any agent.

  • Designed for web software first
  • Model-agnostic
  • Reliability-tested before release
Tooling

Security by default

Sandboxes, permission layers and checks for MCP servers, skills and gateways, built around the risks the audit found.

  • Supply-chain checks (GW-2)
  • Scoped permissions per tool
  • Every action traced
Applied

Enterprise agents

Client work is where each idea is proven. What does not hold up in a real product does not ship to anyone else.

  • Agents in production
  • Findings feed the SDK and tooling
  • A small number of teams at a time

Research tracks05and open questions

Every track starts as a question raised in client work. The answers ship as tooling, not slide decks.

Motiontic research tracks and the question each one works on
TrackThe question it works onRead more
Harness reliabilityHow a long-running agent recovers from tool failures, timeouts and bad model output without a person restarting it.
Agent supply chainWhich frameworks, gateways, MCP servers and skills an agent can trust, and how to check them before every deploy.
Permissions and sandboxingWhat an agent may touch and where its code runs, with access scoped per tool instead of per agent.
Evals and observabilityHow to tell whether an agent got better or worse after a model, prompt or tool change, before users notice.
Agents in web softwareHow agents run inside browser-based products, where few production agent SDKs were designed to live.

Built to pass06the security review

Enterprise buyers ask how an agent's decisions are logged and what happens when it is wrong. Motiontic builds the answers in from the first commit.

  • Evals on every change

    Every prompt or model change runs against a regression suite before it ships.

  • Traces for every decision

    Each agent action can be explained afterwards: what it saw, which tool it called, and why.

  • Human gates

    Refunds, transfers and deletions wait for a person.

  • Data boundaries

    PII masking, scoped permissions and model providers chosen to fit data-residency rules.

  • Audit evidence

    Access controls, change history and policies documented in the shape SOC 2 and ISO auditors expect.

Track record

Startups that sell to fintech giants, through SOC 2, ISO and third-party security reviews.

How Motiontic builds for audits

The companies07behind the logos

From sovereign AI cloud to embedded investing. Eleven companies so far, across AI infrastructure, fintech, developer tools and revenue software.

All11 companies
NeevCloud

NeevCloud

Sovereign AI cloud: GPU and CPU compute with full control over data and workloads.

AI infrastructureVisit NeevCloud
Fring

Fring

An open-source coding agent, priced so individual developers can run it.

Developer toolsVisit Fring
Flo Finance

Flo Finance

Embedded investing and trading: global markets inside any app, through one SDK.

Quantum Heaps

Quantum Heaps

A unified revenue engine for enterprise sales and go-to-market teams.

Revenue & GTMVisit Quantum Heaps
Pulse

Pulse

AI-moderated customer interviews that come back as a themed report with real quotes.

Revenue & GTMVisit Pulse
GM Markets

GM Markets

Tokenized US stocks and ETFs from $1, traded straight from a wallet.

Drevon

Drevon

Evidence-backed prospect research for GTM engineers, with every claim linked to its source.

Revenue & GTMVisit Drevon
Narrative

Narrative

AI testing that helps teams ship quality without hand-written test suites.

Developer toolsVisit Narrative
mirrAR

mirrAR

AR and AI virtual try-on experiences for retail brands.

Retail techVisit mirrAR
Durbin

Durbin

IoT telematics for smarter, safer and more profitable operations.

Quinch Systems

Quinch Systems

AI and IoT solutions, from connected devices to the software that runs on them.

How an08engagement runs

Lab engineers work inside the client's codebase, in hours that overlap the client's time zone, and leave behind a system the team can own.

  1. Day 0

    Call

    Thirty minutes with a lab engineer to decide fit. If Motiontic is the wrong team, the call says so.

    Fit check01
  2. Days 1–5

    Scope

    Motiontic researches the specific problem and writes a scope document with the deliverables, the success measure and the date.

    Clear expectations02
  3. Week 1

    Build

    Engineers join the codebase and Slack and ship production code from the first week.

    In the repo03
  4. Ongoing

    Hand off and support

    Documentation, tests and evals the team can own, followed by continued support after launch.

    Yours to keep04

How the lab09compares

Against building in-house from scratch and hiring a generalist agency. The difference is the harness and the security work underneath.

Motiontic compared with a typical AI development agency and hiring in-house
MotionticBuilding in-house from scratchGeneralist AI agency
HarnessA standard harness, tested across client buildsWritten from zero, made reliable over monthsRebuilt for every project
SecuritySupply-chain checks, scoped permissions and sandboxes by defaultAdded after the first incidentRetrofitted before the audit
Who does the workLab engineers who also run the researchA team already stretched by the roadmapA rotating team behind an account manager
First production codeFirst weekAfter the harness existsAfter weeks of discovery
After launchOngoing support, and every fix feeds the SDKThe team maintains it aloneChange requests
New models and methodsTested in the lab before client workLearned on the company's timeLast year's playbook

Research first,10then client work

New models and methods are tested on Motiontic's own systems before they reach a client. These are the areas the lab works on.

  • 01

    Interaction model

    From clicking steps to supervising outcomes.

  • 02

    AI-native onboarding

    Why empty states cost activation.

  • 03

    Evals

    Golden datasets instead of test cases.

  • 04

    Observability

    Traces, spans and a failure taxonomy.

  • 05

    Cost structure

    Per-action economics, designed in.

  • 06

    The UX of waiting

    Agents that finish while the user is away.

  • 07

    Failure and trust

    Human gates on anything irreversible.

  • 08

    The data layer

    Making what a company has retrievable.

  • 09

    The organisation

    Working agentically after the engagement.

Inside the lab →
Why the lab exists
This may be the most consequential time to build software. The margin between progress and harm is narrow.

AI capability is moving faster than the tooling that keeps it safe. 166 security incidents in one year, in the tools agents are built on, is the clearest sign of that gap.

Motiontic believes growth in AI needs control: agents that are reliable, auditable and safe enough to move people forward. That work is overdue, and it is the lab's focus.

Akash Munshi, founder

Common11questions

What is Motiontic?

Motiontic is an applied AI lab building safer, more reliable AI agents and the tooling around them: a standard harness SDK, security tooling for the agentic stack, and production agents for enterprises.

What is an applied AI lab?

An applied AI lab researches new models and methods and uses them to build production software for real products, rather than publishing research alone. At Motiontic, every technique is tested on the lab's own systems before it reaches client work.

What is an agent harness?

The harness is the code around a model that turns it into an agent: the loop, tool calls, memory, retries, permissions and logging. Most teams write their own from scratch. Motiontic is building a standard harness SDK that can be reused across agents.

What did the Motiontic agentic tooling audit find?

Of 85 agentic tools reviewed, including agent SDKs, harness SDKs, MCP servers, skills and sandboxes, 68 had at least one security incident in 2025–26, for 221 incidents in total, 166 of them in 2026. The most common risk was GW-2, treating frameworks and gateways as critical supply chain: 51 of the 70 tools it applies to had a relevant CVE.

Does Motiontic build agents for clients?

Yes. Client work is where the lab proves its harness and tooling. Engineers work inside the client's codebase and ship production agents, with evals, traces and security controls built in from the first commit.

Can Motiontic build AI that passes SOC 2 and ISO audits?

Yes. Motiontic has helped startups that sell to fintech companies pass SOC 2, ISO and third-party security reviews, with evals, decision traces, access controls and audit evidence built in from the start.

Who does Motiontic work with?

Enterprises and growth-stage companies putting agents into production, especially where an agent touches money, customer data or regulated workflows.

How quickly does work start, and in which time zones?

Engineers join the codebase and ship in the first week after the scope is agreed. The team works remotely across regions and overlaps working hours with clients in the Americas, Europe and Asia-Pacific.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.