New: the Agent Incident Ledger: disclosed incidents of AI gone rogueExplore the ledger

Build vs buysecuring the AI your employees use

AI adoption outpaces the security coverage your team can build in-house. Manifold keeps up with the agents your employees use.

Company-wide adoption of AI agents is barely getting started, and yet security teams already face problems they can’t see or control. Employees try new agents, connect MCP servers and add skills to get work done. Every agent, server and skill has to be found, assessed for risk, and reassessed when an update changes what it can do. That takes focus and investment, and most teams don’t have either to spare.

Building this in-house means your team has to keep up with every new agent, update and tool by hand, for as long as you run it. Manifold covers 100+ agent types and picks up new ones within an hour of first use. Your team can say yes to AI and spend its time on the decisions only it can make: which agents, data and groups are allowed.

85%
of CISOs call safe AI use a top priority
79%
are expected to manage AI risk without more resources or expertise

Your current tools miss what agents do

Endpoint tools see files and commands, and gateways see traffic at one chokepoint, but neither sees an agent’s tool calls and MCP requests, how they connect across a session, or who the agent acted for. Teams end up stitching tools together and building the missing piece themselves. That’s extra work for a team already stretched across endpoints, identity, cloud, data and incidents, whilst the business pushes for more AI.

Promptfrom ravi.menonModel requestapi.anthropic.comTool callread_fileMCP requestjira.searchFile writedeploy.yamlCommandkubectl applyGateway seesEDR seesNot seenwho the agent acted for

Building a prototype is easy. Making it accurate and keeping it current is where the work really starts.

Building a first version has never been faster. Some agents expose hooks or logs, and a capable engineer with a coding agent can capture tool calls from one agent on one laptop in a few days.

Version one shipsEvery month after1New agents2Updates3Supply chain4Browser and cloud5Risk and threat model decisions1New agents2And again

The work starts after that first version ships:

  1. 01New agents keep arriving.

    Employees adopt new tools every month, often before anyone adds them to a list. Each one needs its own integration.

  2. 02Agents keep changing.

    Every update can move a config file, rename an event or add a capability your build doesn’t see.

  3. 03The supply chain grows.

    MCP servers, skills and extensions change daily, and each needs a verdict before an agent relies on it.

  4. 04Agents don’t stay on the laptop.

    The same employee uses AI in the browser and in the cloud, so a laptop-only build leaves blind spots.

  5. 05Logs aren’t decisions.

    Seeing a tool call is the start. Ranking the risk, linking it to a person and responding without stopping the work is the hard part.

FROM THE FIELD
One enterprise security team spent two months building shadow AI visibility from browser logs. It covered one AI tool and missed every coding agent their developers run in the terminal. They then chose Manifold.

What an in-house build has to cover to secure agents

A build that a security team can rely on needs all of the following, and needs to keep them current as agents change.

  1. Discovery

    Find every agent employees use, including shadow AI, on endpoints, in browsers, in VMs and in the cloud.

    ONGOING
  2. Supply chain verdicts

    Judge every MCP server, skill and extension before an agent relies on it.

    ONGOING
  3. Risk assessment

    Rate how risky each agent is based on how it’s set up: what it can access, what it can do without asking, and which combinations could let data leak out.

    ONGOING
  4. Runtime visibility

    Link each tool call, MCP request, file access and command to the person the agent acted for.

    ONGOING
  5. Threat detection

    Analyze the full session, so attacks that unfold across several steps are caught even when each step looks harmless.

    ONGOING
  6. Responses beyond allow or block

    A hard block on a false positive stops real work, so a build needs lighter options too: alerting the security team, asking the developer to confirm, advising the agent on a safer path, or correcting the action before it runs.

    ONGOING
  7. Policy at scale

    Set rules once, tie them to identity provider groups and evaluate them on every action.

    ONGOING
  8. Integrations

    Send findings to your SIEM and Slack with the context to act on them.

    ONGOING
  9. Fleet deployment

    Roll the collector out to every machine and keep it updated.

    ONGOING

Building this in-house means a dedicated engineering team, pulled off other security work, because the agents never stop changing.

Three paths compared

The build estimate usually prices version one. The real cost is every month after it, and what your team has to own to keep it working depends on the path you take.

  • Your team builds and maintains it
  • Partly covered
  • Not covered
  • Manifold runs it
Ongoing work
Full in-house buildyour own platform
Agent hooks plus your existing toolshooks into your EDR or SIEM
Manifoldruns it for you
Covering new agents
Full in-house buildYou build and maintain it: An engineering ticket each time an agent appears or updates
Agent hooks plus your existing toolsPartly covered: Only agents that offer hooks or logs
ManifoldCovered: 100+ agent types covered. New agents picked up within an hour of first use
Where it works
Full in-house buildYou build and maintain it: Wherever you deploy it
Agent hooks plus your existing toolsPartly covered: Laptops, plus cloud agents if set up centrally. Not browser AI or AI in Slack
ManifoldCovered: Endpoints, browsers, VMs and the cloud, from one platform
Supply chain
Full in-house buildYou build and maintain it: Your team picks sources and maintains verdicts
Agent hooks plus your existing toolsNot covered: Not covered
ManifoldCovered: Verdicts from Manifest research across 14 open-source sources
Risk models
Full in-house buildYou build and maintain it: Your team writes and tunes each one
Agent hooks plus your existing toolsNot covered: Not covered
ManifoldCovered: Dozens of managed risk profiles, ranked by owner, context and scale
Runtime view
Full in-house buildYou build and maintain it: Built and maintained by your team
Agent hooks plus your existing toolsPartly covered: Raw events, stitched together in your SIEM
ManifoldCovered: One session view linking the user, agent, tool calls and actions
Threat detection
Full in-house buildYou build and maintain it: A standing detection engineering function
Agent hooks plus your existing toolsYou build and maintain it: Your team writes detections in the SIEM
ManifoldCovered: Dozens of out-of-the-box threat detections that analyze the full session
Responses
Full in-house buildPartly covered: Log or block, unless you build more
Agent hooks plus your existing toolsPartly covered: Allow or block, with separate rules written for each agent
ManifoldCovered: Six responses: Alert, Ask, Advise, Correct, Block call and Block session
Deployment
Full in-house buildYou build and maintain it: Build, ship and update your own collector
Agent hooks plus your existing toolsYou build and maintain it: Deploy through MDM, with a separate hook setup for each agent
ManifoldCovered: Lightweight collector, up and running in minutes through your existing MDM, no proxy
Staffing
Full in-house buildYou build and maintain it: A permanent team to run and maintain it
Agent hooks plus your existing toolsYou build and maintain it: Your team, per agent
ManifoldCovered: Your policy owners

Some teams start from browser logs or EDR data instead. Those see less than hooks: browser logs miss agents in the terminal, and EDR sees files and commands but not the agent’s tool calls. Adding any of them to your SIEM gives you more events, but not one view of what each agent did and who it acted for.

Questions to ask before you build

  • Is AI agent security part of what your company sells, or just something it needs? Most companies build what sets them apart and buy the rest.
  • How many agents do employees use today, and how many will they use in a year?
  • Can you staff a team to maintain this, on call, for the next three years?
  • What happens when a rule fires on legitimate work? Can you respond without slowing the work down?
  • While the build is in progress, which agents run with no one watching?
  • Can you show the board a ranked view of AI risk by owner and business group?

What you keep, what Manifold runs

Buying the platform doesn’t mean handing over control. The decisions stay with your team.

Your team keepsManifold runs
Which agents, groups and data are allowedDiscovery of every agent and its supply chain
Your risk tolerance and exceptionsRisk profiles and supply chain verdicts
Policies, managed as code if you preferEvaluating those policies at runtime on every action
Your identity groups in Okta or Microsoft Entra IDLinking every action to the person the agent acted for
Your findings, in your SIEM and SlackThreat detections and keeping up with new agents
  • Your team keepsWhich agents, groups and data are allowed

    Manifold runsDiscovery of every agent and its supply chain

  • Your team keepsYour risk tolerance and exceptions

    Manifold runsRisk profiles and supply chain verdicts

  • Your team keepsPolicies, managed as code if you prefer

    Manifold runsEvaluating those policies at runtime on every action

  • Your team keepsYour identity groups in Okta or Microsoft Entra ID

    Manifold runsLinking every action to the person the agent acted for

  • Your team keepsYour findings, in your SIEM and Slack

    Manifold runsThreat detections and keeping up with new agents

Because policies follow your identity provider groups and findings land in tools you already run, your program isn’t locked inside one console.

Frequently asked questions

  • Should we build or buy security for the AI our employees use?

    Building a first version is easy. Keeping up is the hard part: every new agent, update and supply chain component needs coverage, which is a permanent workload. Most teams buy the platform that keeps up and spend their own time on the decisions only they can make.

  • Can’t we use the hooks agents already provide?

    Hooks help for the agents that offer them, and some also work in the cloud if an admin sets them up centrally. They don’t cover browser AI or AI in Slack, and each agent needs its own rules. Your SIEM only sees what happened after the fact, so you get events, not risk rankings or responses beyond allow or block.

  • Our EDR already runs on every laptop. Isn’t that enough?

    EDR sees the files and commands on a laptop, but not the agent’s tool calls and MCP requests, how they connect across a session, or who the agent acted for. Manifold looks inside each agent’s runtime and links every action to the person the agent acted for.

  • How long does Manifold take to deploy?

    The lightweight collector is up and running in minutes through your existing MDM, with no proxy and no traffic rerouting. Once it’s running, new agents are picked up within an hour of first use.

  • Will we be locked in?

    Policies attach to your identity provider groups and can be managed as code. Findings go to your SIEM and Slack, so the record of your program lives in your own tools.

  • Will Manifold slow our people down?

    No detection rule is perfect. Ask, Advise and Correct keep work moving where a rule may be too strict, and hard blocks are kept for the clear-cut cases.

Spend your team’s time on policy, not plumbing.

Manifold gives security teams visibility and control over all the AI their employees use. Your team sets the rules, and Manifold keeps up with the agents.

See Manifold in action.

A short tour of the platform, from the first agent discovered to the first policy enforced.