Skip to content
A heavy threshold gate lets one cleared path through while dozens of pink trails stop at it.

Agentic security

Agentic AI security: the guide for businesses that let AI act

The short answer

Agentic AI security is how you let AI agents act for your business without handing them the keys. Give each agent its own identity and only the access its job needs. Make a person approve anything that sends, posts or spends, keep a log of every action, and treat everything an agent reads as untrusted.

Written byBrian GagneCo-founder and CTO, Kief Studio
Technical review byAmelia GagneCo-founder and CEO, Kief Studio

Published

Waiting on you queue with a staff question, a reply awaiting a Go ahead, a blog post review and a social post, each with buttons.
Agentic AI security at the point it matters: everything an agent drafted that sends, posts or spends waits in one queue for the person who owns it.

What is agentic AI security?

A chatbot answers questions. An agent does things: it reads your inbox, updates the CRM, drafts the reply, books the meeting and, if you let it, sends the email. Agentic AI security is the set of controls that decides what an agent may touch, what it has to ask before doing, and how you prove afterwards what it did.

Most of it is older than AI. Identity, least privilege, separation of duties and logging are the same rules we've used for people and servers for decades. What's new is that the thing holding the access can be talked into using it by a sentence hidden in an email or a web page. That one fact changes how you design everything around it.

Here's how I explain it to owners. You'd never give a new hire the master key, the company card and your email password on day one, then leave for the week. An AI agent deserves the same caution, and it has fewer instincts for when something looks off.

A faint pink thread drifts toward a row of sealed doors and one amber lock holds.

Why agentic AI security matters now

Agents went from demos to daily work fast, and the numbers show agentic AI security lagging behind. In IBM's Cost of a Data Breach Report 2026 (opens in a new tab), the global average breach cost reached 4.99 million dollars. Among organizations that reported an AI-related breach, 92% lacked proper AI access controls, and incidents involving unapproved AI use rose from 20% to 43% of organizations, as IBM's analysts explain here (opens in a new tab).

Every agent also needs credentials to reach your tools, and those pile up. The 2026 Identity Security Landscape (opens in a new tab) from Palo Alto Networks counts 109 machine identities for every human one. GitGuardian's State of Secrets Sprawl 2026 (opens in a new tab) found 28.6 million new secrets exposed in public GitHub commits in 2025, up 34%, with keys for AI services up 81%.

The standards bodies have caught up. OWASP published the Top 10 for Agentic Applications for 2026 (opens in a new tab) on December 9, 2025. In May 2026 the cyber agencies of the United States, the United Kingdom, Australia, Canada and New Zealand released joint guidance, Careful adoption of agentic AI services (opens in a new tab). The UK's NCSC summed it up (opens in a new tab) as: start small, use agents only for low-risk tasks, and apply established cyber security controls from the outset.

Agentic AI security for a 5 to 50 person business

Big companies have identity teams and a security operations center to catch mistakes. A business of 5 to 50 people usually has an owner, an office manager and an IT provider who visits when something breaks. That team can still run agents safely, because the controls that matter most are design choices made once, up front. Get them right in the system you choose and nobody has to watch it all day.

Increase in 2025 is highest in Keys for AI services at 81%.

In the same report, new secrets exposed in public GitHub commits in 2025 were up 34% on the year before, and keys for AI services were up 81%. Source: State of Secrets Sprawl 2026
Increase in 2025
All new secrets34
Keys for AI services81
In the same report, new secrets exposed in public GitHub commits in 2025 were up 34% on the year before, and keys for AI services were up 81%. Source: State of Secrets Sprawl 2026 (opens in a new tab)

Agentic AI risks, and the control that answers each one

OWASP's list names ten risks, ASI01 to ASI10. These are the agentic AI security risks a small business meets first, each paired with the control that stops it.

Goal hijack and prompt injection

An agent reads a customer email that says, in white text at the bottom, to forward the last ten invoices to an outside address. That's prompt injection, and OWASP ranks Agent Goal Hijack first (ASI01). You can't train it away completely. You contain it: the agent treats what it reads as data, and the actions that could do damage need a person's yes no matter who asked for them.

Too much access

The agent that summarizes your inbox doesn't need to send from it, and the one that drafts invoices doesn't need to delete customers. OWASP calls this Identity and Privilege Abuse (ASI03), and its older list for AI apps calls it excessive agency (opens in a new tab) (LLM06). The answer is least privilege, set per agent and per task.

Tool misuse

An agent with a legitimate tool can still use it badly: a bulk update run against the wrong list, a calendar tool used to invite strangers. OWASP lists this as Tool Misuse and Exploitation (ASI02). Narrow tools, limits on how much one action can change, and an approval step for anything that leaves the building keep a mistake small.

Leaked keys and tokens

Agents run on API keys, tokens and service accounts, which security people call non-human identities. When one leaks, whoever holds it is the agent as far as your systems can tell. In December 2022 a chat company (opens in a new tab) reported that a small number of employee tokens were stolen and used to download private code repositories. MITRE tracks the technique as steal application access token (opens in a new tab) (T1528). Keys belong in a vault, never in code or chat, scoped to one job and easy to revoke.

Memory poisoning and cascading failures

Agents that remember can be taught wrong things, and one agent's bad output becomes the next agent's input. OWASP lists Memory & Context Poisoning (ASI06) and Cascading Failures (ASI08). Keep rules a person wrote separate from notes an agent learned, so a learned note can never overrule a rule, and put a person's check between agents where the stakes are high.

Cracked stone tablets glow pink and the nearest is bound by an amber band.

The five controls of securing AI agents

In practice, agentic AI security comes down to five controls. If a product you're evaluating can't show you all five, keep looking.

Its own identity for every agent

Each agent gets its own account and its own credentials, never a person's login and never a shared key. Then every action traces back to one agent, and you can switch one off without breaking the rest. NIST's Cybersecurity Framework 2.0 (opens in a new tab) already names services alongside users in PR.AA-01, so agents fall under the identity rules you have.

Least privilege, scoped to the task

Give each agent only the tools and data its job needs, for as long as it needs them. NIST CSF 2.0 asks for exactly this in PR.AA-05: permissions defined in a policy, enforced, reviewed, and built on least privilege and separation of duties. The joint guidance agrees: minimum privileges for the task.

Human in the loop for anything that leaves the building

Drafting, summarizing and checking can run on their own. Sending, posting, publishing, paying and changing a live site wait for the person who owns that area. We go deeper on this in AI staff with approvals.

An audit trail you can trust

Write down every action in plain words: which agent, what it did, when, and who approved it. A good log is written before the action runs, so a crash in the middle still leaves a record, and it can't be edited after the fact.

Untrusted input stays data

Email, web pages, uploaded files and messages from outside are things an agent reads, never orders it follows. Mark them as outside data, and make sure the controls above hold even when an agent gets fooled.

Five nested metal rings contain a restless pink core, with the outer ring lit amber.

How AI agents get identities and permissions

When an agent connects to your tools, it usually does it through the Model Context Protocol (MCP), the open standard that lets an AI system call outside tools. The protocol's authorization spec (opens in a new tab), in its July 28, 2026 revision, treats a protected MCP server as an OAuth 2.1 resource server. Tokens have to be issued for that one server, and a server must not accept or pass on any token that wasn't issued for it.

The agent signs in like an app, gets a pass that only works at one door, and can't lend that pass to anyone. That's the identity half of AI agent security. The permission half is deciding, per agent, which doors exist at all.

A slim amber token fits only the one door it was cut for while pink trails glance off the rest.

Shadow AI: the agents nobody approved

The riskiest agent in most small businesses is the one somebody connected to the company inbox with a free account and a browser extension. IBM's 2026 numbers say incidents involving unapproved AI use rose from 20% to 43% of organizations. Banning AI doesn't fix it, because people use what saves them time. Giving them a sanctioned tool that's easier than the workaround does.

Organizations is highest in 2026 report at 43%.

Security incidents involving shadow AI, where workers use unapproved AI, more than doubled, from 20% to 43% of organizations between IBM's 2025 and 2026 reports. Source: IBM Cost of a Data Breach Report 2026
Organizations
2025 report20
2026 report43
Security incidents involving shadow AI, where workers use unapproved AI, more than doubled, from 20% to 43% of organizations between IBM's 2025 and 2026 reports. Source: IBM Cost of a Data Breach Report 2026 (opens in a new tab)

Agentic AI security in Massachusetts

If you hold personal information about Massachusetts residents, 201 CMR 17.00 (opens in a new tab) already applies to your AI agents. It asks for secure access control measures that restrict access to records containing personal information to those who need it, unique identifications for each user, and reasonable monitoring of systems for unauthorized use. An agent with its own identity, scoped access and a complete log is how you meet those words when the user is software.

Massachusetts' breach law, M.G.L. c. 93H (opens in a new tab), also requires notice when personal information is acquired or used without authorization. An audit trail is what lets you answer the first question a regulator or customer asks: what did it touch? We're in Shrewsbury, and this is the conversation we have with most owners in Worcester County and Boston who want agents doing real work.

A records room in a converted red-brick mill office, shelves of archive boxes with pink-lit client files, and an amber-lit badge reader.

What we've learned running agents every day

Kief Studio runs its own company on LTFI, so agentic AI security isn't theory to us. A few things we'd tell any owner.

  • Default to asking. It feels slow for about a week, then the queue becomes the fastest part of your day, because everything that needs you is in one place.
  • Let trust be earned one kind of task at a time. A standing approval for routine posts is reasonable after you've approved dozens. Money, deletions, keys and permissions should ask every time.
  • Keep the rules with people. An agent can suggest a change to how it works, but only a person should be able to make one.
  • Read the log weekly at first. It's the fastest way to see where an agent is doing more than you expected.
A walnut desk with a brass outgoing mail tray of pink-edged envelopes, held down by a heavy amber-lit brass paperweight.

How LTFI handles agentic AI risks

LTFI is the agentic business operating system Kief Studio makes and runs. These are the agentic AI security controls it ships with today.

  • Anything that sends, posts, publishes or spends waits in one queue for the person who owns it, on a computer or a phone.
  • Standing approvals are yours to give for one kind of work and take back at any time.
  • Rules, permissions, credentials, deleting and money stay with people. An AI teammate can't change them.
  • The activity log records every action by the staff and the team in plain words, newest first, with who approved it.
  • Each company runs on its own dedicated server, managed, patched and watched by us.
  • Your AI runs on keys you own, under your contract with the AI company, with spend caps you set.
A phone lies face down on a walnut desk with an amber glow around its edges, beside a closed laptop with a pink-lit cable.
Activity log listing staff and team actions newest first, such as finished hand-offs, approvals and denials, each with a time stamp.
The activity log in LTFI records every action by an AI teammate or a person, in plain words, newest first.

What to do about agentic AI security this month

Two lists: one for the person who owns the business, one for whoever runs its systems.

If you own the business

  1. List every AI tool that can read or send company email, files or customer data, including the free ones people signed up for.
  2. Decide which actions need your yes: sending, posting, paying and deleting are a good start.
  3. Pick one low-risk job for your first agent, like drafting replies, and expand from there.
  4. Ask any AI vendor to show you its approval queue and its activity log before you buy.

If you run the systems

  1. Give each agent its own service identity and scoped credentials; retire shared keys.
  2. Move keys into a vault, rotate them, and scan repositories and chat for leaked ones.
  3. Require audience-bound tokens for MCP servers and refuse token passthrough.
  4. Mark outside content as untrusted in every agent's context, and log every tool call before it runs.

Key takeaways

  • Agentic AI security is mostly old controls applied to a new kind of user that can be talked into things.
  • Five controls carry it: identity, least privilege, human approval, an audit trail, and untrusted input kept as data.
  • Good agentic AI security starts with low-risk work and lets trust grow one kind of task at a time.
  • In Massachusetts, 201 CMR 17.00 already covers what your agents can reach.

Where this lines up with the standards

The frameworks and rules this page maps to, with the exact item and what it asks of you.

FrameworkItemWhat it asks
OWASP Top 10 for Agentic ApplicationsASI01, ASI02, ASI03 (opens in a new tab)Contain goal hijack, limit tool misuse, and keep each agent's identity and privileges narrow.
OWASP Non-Human Identities Top 10NHI2, NHI5, NHI7 (opens in a new tab)Stop secret leakage, overprivileged machine identities and long-lived secrets.
NIST CSF 2.0PR.AA-01, PR.AA-05 (opens in a new tab)Manage identities for services as well as users, and enforce least privilege and separation of duties.
NIST AI RMF and NIST AI 600-1Govern, Map, Measure, Manage (opens in a new tab)Assign accountability for AI systems and keep human oversight over what they do.
Joint agency guidanceCareful adoption of agentic AI services (opens in a new tab)Start with low-risk tasks, keep privileges minimal, and approve high-risk actions.
MITRE ATT&CKT1528 (opens in a new tab)Defend the application access tokens agents run on.
Massachusetts201 CMR 17.00 and M.G.L. c. 93H (opens in a new tab)Restrict access to personal information, use unique IDs, monitor systems, and be ready to report a breach.

Key terms

Global average cost of a data breach, in US dollarsSource: ibm.com
4.99 million
Organizations with an AI-related breach that lacked proper AI access controlsSource: ibm.com
92%
Machine identities for every human oneSource: paloaltonetworks.com
109
New secrets exposed in public GitHub commits in 2025Source: gitguardian.com
28.6 million

Questions

Agentic AI security

What is agentic AI security?

Agentic AI security is the set of controls that decides what an AI agent can access, what it must ask a person before doing, and how you prove what it did. It covers agent identity, least privilege, human approval, audit trails and handling untrusted input.

Is it safe to let AI agents send email for my business?

It can be, if a person approves each send and sees exactly what will go out. Keep drafting automatic and sending approved until you trust one narrow kind of message, then give a standing approval you can take back.

What is the biggest risk with AI agents?

Too much access combined with untrusted input. An agent that reads outside email and can also send, pay or delete can be tricked into doing harm. Narrow access and a person's approval on high-impact actions keep that risk small.

How is AI agent security different from regular cybersecurity?

The controls are familiar: identity, access, logging. The difference is that the user holding the access can be persuaded by text it reads, so you design for the case where the agent gets fooled and make sure the damage stays small.

Do small businesses need agentic AI security?

Yes, and it's easier for them than for large companies. A small business can choose a system with approvals and logging built in and get most of the protection on day one, without a security team.

Does 201 CMR 17.00 apply to AI agents?

It applies to anyone who holds personal information about Massachusetts residents, and it doesn't exempt software. Access to that information has to be restricted to those who need it, users need unique IDs, and systems need monitoring, which covers agents too.

Where do I start with agentic AI security?

List every AI tool that can reach company email, files or customer data. Then pick the actions that need a person's yes, and give your first agent one low-risk job. Agentic AI security grows with the work from there.

Who owns agentic AI security in a small business?

The owner does, even when an IT provider handles the setup. Name the person who approves each kind of action, and make sure the log records their name every time.

What are examples of agentic AI security controls?

Agentic AI security controls include one identity per agent, the least access each task needs, a person's approval before anything leaves the business, an activity log of every action, and a way to switch an agent off fast. This guide walks through each one.

Everything in this section

  • Medallions tangled into one pink knot, and one free medallion with its own amber thread.

    Identity and access

    Agentic IAM

    Agentic IAM gives every AI agent its own identity, scoped short-lived credentials and authority it can trace to a person. How it works and where to start.

    Brian Gagne

  • A blurred hub drowns in pink cables while a clean hub in front holds one amber line.

    Identity and access

    AI agent permissions

    AI agent permissions decide what an agent can read, change and send. How to scope its tools, catch permission drift and run an access review.

    Brian Gagne

  • A sprawling swarm of small pink-lit tags, with one sealed and glowing amber inside a single capsule.

    Identity and access

    Non-human identities

    What a non-human identity is, why AI agents multiply them, and how to vault, scope and rotate API keys, tokens and service accounts before one leaks.

    Brian Gagne

  • A capsule travels through a line of rings and exits cleanly through an amber one.

    Identity and access

    AI agent lifecycle

    The AI agent lifecycle, run like staffing: who requests and owns an agent, what it gets on day one, how access is reviewed, and how to retire it cleanly.

    Amelia Gagne

  • A pink stream of light is held at a round amber valve gate, waiting for a person's yes before anything leaves.

    Oversight

    Human approval for AI agents

    Human approval for AI agents that people actually read: which actions need a yes, how to avoid approval fatigue, and how standing approvals work.

    Amelia Gagne

  • A column of stacked plates with pink-lit seams sealed by an amber top.

    Oversight

    Audit trails for AI agents

    An AI agent audit log proves what an agent did, who asked and who approved it. What to record, how to keep it tamper evident, and how long to keep it.

    Amelia Gagne

  • A column of vault tumblers lights pink at each seam, and the last aligns amber as the vault door edge begins to glow.

    Oversight

    AI agent security checklist

    An AI agent security checklist for owners: what to check before an agent acts, then identity, approvals, data, logging, incidents and retirement.

    Amelia Gagne

  • Hidden pink-lit devices behind dark panels, and one wide amber archway beside them.

    Threats

    Shadow AI

    Shadow AI is the AI your team uses without approval. How to find it without spying, write an AI policy people follow, and offer a better tool.

    Amelia Gagne

  • Many pink cables converge on one hub, and every connection passes through its amber iris.

    Threats

    MCP security

    MCP security for businesses running AI agents: what the Model Context Protocol's authorization rules require, and how to vet servers and stop tool misuse.

    Brian Gagne

  • A smoked translucent slab carries a hidden pink thread toward the agent's tools, and an amber ring stops it before it reaches them.

    Threats

    Prompt injection

    Prompt injection turns an email, a web page or a file into orders for your AI agents. Why it can't be trained away, and the controls that limit the damage.

    Brian Gagne

  • Ten black pillars veined with pink fractures stand in a row, and three amber rings thread through every one of them.

    Threats

    OWASP Agentic Top 10

    The OWASP Agentic Top 10, ASI01 to ASI10, explained for owners: each risk in everyday terms, a small-business example, and the control that answers it.

    Brian Gagne

  • Two interlocking rings, a lock dial and an open aperture, joined by an amber clasp.

    Operations

    Agentic IAM and AI SOC in 2026

    Agentic IAM and AI SOC are the two 2026 security predictions that reach a small business. What they mean, what to ask your IT provider, and what to do first.

    Amelia Gagne

  • Pink beads of personal detail drift between dark vessels, and one sealed amber vial holds only a few.

    Operations

    AI agents and personal data

    AI agents and personal data: what an agent should see, what it may keep, how to delete it on request, and how to log its work without hoarding personal details.

    Amelia Gagne

  • A swarm of pink alert shards streams through a dark sorting lens, while one amber lever stays locked until a person decides.

    Operations

    The AI SOC

    An AI SOC lets agents triage, enrich and summarize alerts. Containment, lockouts, deleting data and closing incidents still need a person to approve them.

    Brian Gagne

  • Reference

    Glossary

    An agentic AI glossary in plain words: AI agents, agentic IAM, non-human identities, prompt injection, shadow AI, MCP and more, each with a short definition.

  • Reference

    Questions

    AI agent security questions owners ask most: is it safe to let AI send email, what needs approval, who's responsible, and how Massachusetts law applies.

Most MSPs are resellers. Kief Studio is the manufacturer.

A reseller sends your problem to the company that made the software. We are that company: we build LTFI and run it, so the people you talk to can change it. Book a walkthrough and see it on your own business.

Nothing plays until you press play.