AI agent security risks: what can go wrong at work
An agent does not just say the wrong thing. It can do the wrong thing, with a real employee's sessions, files and tokens. Here is what goes wrong, how it happened in real cases, and which controls stop it.
What raises an agent's risk
Relative weighting for illustration, not measured data.
Next step
AI Agent Blocklist: keep high-risk agents off your network and devices
Every agent domain and device identifier carries a risk grade, so you can block the riskiest agents first.
From $99 a month for one organisation
Are AI agents a security risk? The short answer
Yes, in a specific way. The model is rarely the weak point. The weak point is the combination of access, untrusted input and the ability to act, inside a company that does not know which agents are running.
They hold real access
Agents use the employee's browser sessions, OAuth grants, API keys and files. Whatever the user can do, the agent can do.
They read untrusted text
Web pages, mail, tickets and documents feed the agent's next step. Some of that text is written by attackers.
They act at speed
An agent can send, upload, buy or delete in seconds, many times over, before a person looks at the result.
The ten AI agent security risks admins meet most
In December 2025 the OWASP GenAI Security Project published a Top 10 for agentic applications, led by agent behaviour hijacking, tool misuse and exploitation, and identity and privilege abuse.
Our list below follows the same ground, sorted by what admins actually see on networks and devices.
Redirected goals High
Hidden instructions in a page, email or file replace the user's task with the attacker's.
Bites in: AI browsers, agent extensions, mail agentsTool misuse High
The agent uses a legitimate tool in a harmful way: a send function, a file write, a payment form.
Bites in: agents with connectors and MCP toolsBorrowed identity and privilege High
Agents act as the user with every right the user has. Logs show a person, not an agent.
Bites in: browser and desktop agentsData out through actions High
Data leaves through normal actions: a reply, an upload, a form, a URL with data in it.
Bites in: any agent that can write or browseUntrusted supply chain High
Harmful extensions, skills, packages and MCP servers can look like useful add-ons.
Bites in: developer machines, extension storesUnexpected code execution High
Coding and desktop agents run commands. A bad plan becomes a deleted folder or a wiped database.
Bites in: coding agents, personal agentsPlanted memory Medium
Agents that remember can store a planted instruction and act on it days later.
Bites in: assistants with long-term memoryCascading errors Medium
One agent's wrong output becomes the next agent's input. Mistakes grow along the chain.
Bites in: multi-agent workflowsApproval fatigue Medium
People click Allow on the tenth prompt without reading it. Confirmations stop protecting anyone.
Bites in: agents with many permission promptsSprawl and no owner Medium
Nobody knows how many agents run or who answers for them. Old tokens stay valid.
Bites in: every organisation without an inventoryThe combination that lets data leave through an agent
Any one of these is normal. An agent that has all three can be talked into sending data out by a single hidden sentence, with no malware involved.
- Private data: your mailbox, drive, CRM, code or HR system
- Untrusted content: public web pages, inbound mail, shared documents, issue trackers
- A way out: sending mail, posting comments, uploading files, loading a URL
Remove one circle and the attack gets much harder. Blocking unapproved agents removes all three at once.
Patterns reported in 2025 and 2026, and the lesson in each
These patterns were widely discussed in 2025 and 2026. They are not tied to one product: any agent with the same access can show them. Each comes with the lesson for admins.
- 2025
Coding agents with write access to live systems removed data that their users had told them to leave alone. Instructions in a chat did not hold the agent back.
Lesson: agents with write access to production need hard limits, not instructions. - 2025
Researchers showed that hidden text on a public page can steer an AI browser's assistant into opening the user's mailbox and sending a one-time login code to an outside address.
Lesson: an agent inside a logged-in browser can cross from one site to another on its own. - 2025
Lookalike MCP server packages, published by people unrelated to the service they named, can change behaviour in a later version, for example by copying outgoing messages to an outside address.
Lesson: MCP servers are software supply chain. Allow only reviewed ones. - December 2025
Gartner published advice titled "Cybersecurity Must Block AI Browsers for Now", and OWASP released its Top 10 for agentic applications.
Lesson: blocking AI browsers is now mainstream security advice. - 2026
Researchers pointed out that public add-on stores for self-hosted agents accept skills from anyone, and that agent control panels are often left reachable from the internet.
Lesson: self-hosted personal agents bring their own add-on store and their own services to manage. - 2026
Researchers showed that a disguised instruction in an email can lead a cloud agent to run code and reach stored credentials for connected services, when the user only asked it to check the inbox.
Lesson: filters that catch plain commands miss disguised ones. Limit what agents can reach.
Operating system makers now say the same. Microsoft's documentation for agent features in Windows names cross-prompt injection, where content in a document or interface overrides an agent's instructions, as a risk that can lead to data leaving the device or unwanted software being installed.
Risk by agent type
Not every agent carries the same risk. Three dots means the factor is usually strong for that type; check each tool for its own settings.
| Agent type | Untrusted input | Data reach | Can act outside | Hard to see | Read more |
|---|---|---|---|---|---|
| AI browsers | ●●● | ●●● | ●●● | ●●● | AI browser risks |
| Agent extensions | ●●● | ●●● | ●●● | ●●● | Are AI extensions safe? |
| Computer-use agents | ●●● | ●●● | ●●● | ●●● | Computer use agents |
| Coding agents | ●●● | ●●● | ●●● | ●●● | Coding agent risks |
| MCP servers | ●●● | ●●● | ●●● | ●●● | MCP security risks |
| Personal agents | ●●● | ●●● | ●●● | ●●● | OpenClaw risks |
| Hosted agent platforms | ●●● | ●●● | ●●● | ●●● | Category in our list |
Two patterns stand out. Agents that live inside the browser score high everywhere, because the browser already holds every session. And agents that run on a developer's machine combine source code, cloud keys and a shell, so one bad command reaches far.
Hosted platforms look safer because nothing is installed, but they hold OAuth grants to your SaaS apps and keep working when the laptop is closed.
Who is most exposed in a typical company
Risk follows access. The people who install agents most eagerly are often the ones whose accounts can do the most damage.
| Group | Typical agent | What an agent could reach |
|---|---|---|
| Executives and their assistants | Mail and calendar agents, AI browsers | Board papers, deal mail, travel and payment details |
| Finance and procurement | Browser agents filling forms, invoice agents | Banking portals, supplier records, approvals |
| Developers and IT | Coding agents, MCP servers, personal agents | Source code, cloud keys, admin consoles, shell |
| Sales and support | Agent extensions in CRM and help desk tabs | Customer records, exports, outbound mail |
| HR and school staff | Note and summary agents, browser assistants | Staff, pupil and patient records |
Agentic browsers, multi-agent chains and agents that spend money
Agentic browser security risks
The agent sits where your users are already signed in to everything. One page with planted text can reach every other tab. See prompt injection in AI browsers and Gartner's advice to block AI browsers.
Multi-agent risks
Agents hand tasks to other agents and trust their output. A misled or confused agent passes its error downstream, and tracing the cause gets hard.
Agents that buy and pay
Agents that can check out, subscribe or call metered APIs turn mistakes into invoices. Wrong orders and runaway loops cost real money.
Autonomy multiplies everything above. An agent that asks before each step limits damage to one step. An autonomous agent that runs for hours in the background can repeat a bad action many times before anyone sees it.
AI agent risks and benefits: keep both in view
Why teams want agents
- Repetitive web and data entry work done in minutes
- Research and summaries across many sources
- Code changes and tests prepared for review
- Reports, tickets and follow-ups drafted on a schedule
What makes that safe enough
- A short list of reviewed agents, each with an owner
- Least access: dedicated accounts, scoped tokens, allowed sites
- Confirmation before sending, buying and deleting
- Everything else blocked on the network and on devices
Many organisations land on exactly this split. Our page on why companies block AI agents collects the reasons they give.
Signs an agent is already causing trouble
Agent incidents rarely announce themselves. They look like odd behaviour from a normal user account. Watch for these patterns in the logs you already keep.
Mail at machine speed
A user account sends many messages or forwards within seconds, often to addresses it never wrote to before.
Bulk reads, then uploads
Large numbers of files opened in a short burst, followed by uploads to a site the team does not normally use.
New grants with wide scopes
Fresh OAuth consents for apps that ask to read and send mail, or to read every file in a drive.
Activity outside working hours
Sessions active at night or at weekends from a laptop that should be closed: a scheduled agent at work.
Shells started by unusual parents
Command shells launched by an editor plugin, a browser helper or an unfamiliar background app.
Agent domains in DNS
Repeated lookups to agent platforms, MCP services or AI browser update hosts from devices that should not use them.
Writing agent risks into your risk register
Risk registers need statements an auditor can test. These four cover most agent exposure and point to a control you can show evidence for.
| Risk statement | What makes it likely | Evidence of control |
|---|---|---|
| An unapproved agent acts on company data with an employee's access | No block on agent extensions, AI browsers or desktop agents | Browser and MDM policies with agent identifiers, DNS block logs |
| A harmful add-on copies data or credentials | Open extension stores, open package installs, unreviewed MCP servers | Extension allow list, package rules, MCP allow list |
| An approved agent is steered by planted content | Agent can read untrusted content and send data out | Scoped access, confirmation for outbound actions, allowed site lists |
| Agents run without an owner | No inventory, no review dates | Inventory export with owners and review dates |
How we solve it: controls mapped to each risk
No single control covers every risk. Blocking unapproved agents removes most of the exposure at once; the rest comes from limiting what approved agents can reach.
| Risk | Control | What our list gives you |
|---|---|---|
| Redirected goals, data out | Block AI browsers and agent extensions you have not approved | Agent extension IDs for Chrome, Edge and Firefox, AI browser bundle and package IDs, their domains |
| Untrusted supply chain | Allow only reviewed extensions, packages and MCP servers | MCP server and integration domains, CLI agent package names such as @openai/codex |
| Code execution | Keep coding and personal agents off machines that do not need them | Process names, Homebrew and winget IDs such as OpenAI.Codex |
| Borrowed identity | Separate agent accounts and scoped tokens for approved agents | A clear line between approved agents and everything else |
| Sprawl, no owner | Inventory, owners, review dates | Tool name, category and risk grade per entry for matching |
Four steps that cut most of the exposure this month
- Block by domain. Load the agent domain list into your DNS filter or firewall as a custom or external list. See blocking AI agents at the firewall.
- Block on devices. Push extension block lists to browser policy and app identifiers to MDM.
- Allow a reviewed few. Grade each request with our AI agent risk assessment and give approved agents only the access they need, as set out in agent identity and permissions.
- Stay current. Refresh the files daily. New agents and add-ons appear every day.
Start with the agents you never approved
- 2,600+ AI agent domains and 5,300+ device identifiers
- A risk grade on every entry, so the riskiest go first
- Firewall, DNS, hosts, PAC, CSV, JSON, browser and MDM policy formats
- Updated every day, with access set up within one business day of your order
Full contents: the AI Agent Blocklist.
Ask us or order a plan
Tell us what you need to block and where your rules live. We reply within one business day.
- Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
- Updated every day as new AI browsers, agents and extensions appear
- Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you
AI agent security risks: FAQ
Are AI agents a security risk?
Yes, when they act with real access and nobody controls which ones run. An agent works with the user's sessions, tokens and files, reads untrusted content and can act. A planted instruction or a malicious add-on can turn that access against you.
What are the OWASP AI agent risks?
In December 2025 the OWASP GenAI Security Project published a Top 10 for agentic applications. It covers agent behaviour hijacking, tool misuse and exploitation, identity and privilege abuse, and more.
The rest of the list deals with supply chain, code execution, memory and multi-agent risks. Use the current OWASP text for exact titles.
What are the main agentic AI security concerns?
Prompt injection that redirects the goal, agents acting with borrowed identities and broad permissions, data leaving through normal actions, harmful extensions, skills and MCP servers, and agents nobody owns or monitors.
What are agentic browser security risks?
An AI browser or agent extension acts inside logged-in tabs. Hidden text on a page can steer it to open other sites, read mail or submit forms. Researchers showed this in 2025, and in December 2025 Gartner advised security leaders to block AI browsers for now.
What are multi-agent risks?
When agents pass work to other agents, one bad input or one compromised agent can spread through the chain. Errors compound, trust between agents is assumed, and it becomes hard to see which agent made which decision.
What are the risks of autonomous AI agents?
Autonomous agents run without a person approving each step. The longer they run and the more they can touch, the more a wrong plan, a misread page or a planted instruction can do before anyone notices.
How do we reduce AI agent risks quickly?
Block unapproved agents by domain and by device identifier, allow a short reviewed list, limit what approved agents can reach, require confirmation for sending, buying and deleting, and log agent activity separately where your tools allow it.
OWASP is a trademark of the OWASP Foundation. Gartner is a trademark of Gartner, Inc. Microsoft and Windows are trademarks of Microsoft Corporation. Chrome is a trademark of Google LLC. Firefox is a trademark of the Mozilla Foundation.
Codex is a trademark of OpenAI. OpenClaw is a trademark of its owner. Names are used only to identify the products and organisations discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.