A central issue in AI agent security risks is permissions: a model connected to tools can turn a mistaken instruction into an action. Agents can work through multiple steps toward a goal, but they do not have to act without approval. A deployment may require confirmation before sending email, running code, updating records or placing an order. The risk depends on what the tools can reach and which actions the system lets them execute unattended.
This is not a theoretical worry anymore. In 2026, research from the Cloud Security Alliance and Token Security found that 65% of surveyed organizations reported at least one AI-agent-related security incident in the past year. This guide is the map of what can go wrong when AI acts on its own, so you can spot the risk before you hand an agent the keys.
I got a first-hand reminder of this while writing this post. I let Claude act as an agent in my browser, searching and clicking on my behalf, and before it could do anything it asked for two macOS permissions: full control of my mouse and keyboard, and the ability to record my screen. The access is genuinely useful, and it is also a lot of reach to hand any piece of software. It is exactly the kind of broad access that turns a small agent mistake into a real one.

Letting Claude act in my browser meant granting full mouse and keyboard control plus screen recording. That reach is the point, and the risk.
Tool access turns AI answers into actions
A text-only chatbot returns an answer for someone to act on. A tool-connected assistant or agent can also request actions in other systems. Whether those requests execute immediately depends on its permissions, approval rules and other controls.
That change can turn a wrong answer into a wrong action. If an agent misunderstands a task and its tools permit it, it might email the wrong person, delete the wrong file or approve the wrong refund. Keeping human approval for sensitive actions creates a checkpoint; it does not guarantee that either the model or the reviewer will catch every mistake.
Five AI agent security risks, in plain terms
The failures worth understanding first fall into five groups. OWASP’s agentic list is longer and more granular, and this is the short version rather than the complete one.
1. It does too much (excessive agency)
This is the headline risk, and OWASP lists it as “Excessive Agency.” The model gets something slightly wrong, and because it was allowed to act, the small mistake becomes a real incident: it issues the refund, changes the shipping address, updates the customer record, or runs the command. The danger isn’t that the AI is evil; it’s that it’s confidently wrong while holding real permissions.
2. It gets hijacked by what it reads
Agents may read outside content: web pages, emails, documents, tickets. Hidden instructions in that content can redirect the agent without anyone seeing it, an attack called indirect prompt injection. Because an agent can then act, a hidden line of text can turn into a real action, like leaking data or sending a message. (For how this works, see our guide to direct vs indirect prompt injection.)
3. It exposes or leaks data
Agents may need access to sensitive information, and excessive access increases what a mistake can expose. In the CSA and Token Security 2026 survey, 61% of organizations that reported an agent-related incident experienced data exposure or mishandling. This is a share of affected organizations, not a share of individual incidents. An over-connected agent can pull data from one place and send it somewhere it never should have gone.
4. It uses tools and permissions it shouldn’t
Agents act through tools and credentials. If an agent shares a key with people or other tools, revoking that key affects every user of it. Request-level logs may still distinguish activity, but a separate, narrowly scoped agent credential makes independent revocation and attribution easier.
5. One agent’s mistake spreads to others
As teams chain agents together, a compromised or confused agent can pass bad output to another. If the receiving agent treats that output as a trusted instruction, the error can propagate. Separate permissions and validation at each handoff help limit how far it travels.
AI agent visibility needs evidence
Gravitee’s 2026 report compares stated confidence in agent visibility across its December 2025 and April 2026 survey waves: 82.6% and 91.8%, respectively. The April wave reported mean monitoring coverage of 52%. Those are self-reported survey measures, not an audit proving that every remaining agent has no security controls or that the same agents became less safe.
For a small team, the practical question is whether you can name your agents, identify their permissions and inspect their actions. Confidence alone cannot answer those questions.
In November 2025, Anthropic reported what it described as the first documented large-scale cyberattack executed mostly without substantial human intervention. It estimated that AI performed 80 to 90 percent of the campaign’s activity. Attackers manipulated Claude Code into attempting infiltration of roughly thirty global targets, with success against a small number, according to Anthropic. This was a reported espionage operation, not a measurement of the risk faced by an ordinary small business. It illustrates how a sequence of tool actions can support a harmful goal.
Limit AI agent permissions and damage
You don’t make agents safe by trusting them more carefully; you make them safe by limiting what a mistake can do. Four useful starting points for a small team, and each is its own deeper topic. This is where to start, not the whole of agent security: supply chain, secrets management, network egress and incident recovery all matter once you are past the basics.
- Least privilege. Give an agent access only to the specific tools and data the task needs, nothing more.
- A human in the loop for anything irreversible. Sending money, deleting data, emailing customers: require a person to confirm.
- Its own identity and a log. Give each agent its own credentials, not a shared key, and log what it does, ideally down to which instruction triggered which action, under which permissions, and whether a human approved it. That log tells you what happened and how far it went. It does not undo a sent email or a completed payment, which is why the approval step above sits in front of the actions that cannot be taken back.
- Guard what it reads. Treat outside content as untrusted, because it can carry instructions intended to hijack the agent.
Each of these is expanded in the rest of this series, from limiting agent permissions to understanding the tools and connectors agents rely on. Start here, then go deeper where it matters for you.
FAQ
QCan an AI agent take actions without my approval?
Can an AI agent take actions without my approval?
QAre AI agents safe to connect to my email, files, and accounts?
Are AI agents safe to connect to my email, files, and accounts?
QWhat is the most common way AI agents cause security incidents?
What is the most common way AI agents cause security incidents?
QCan I trust an AI agent with payments or other sensitive actions?
Can I trust an AI agent with payments or other sensitive actions?
QHow do I stop an AI agent from doing something I didn’t intend?
How do I stop an AI agent from doing something I didn’t intend?
Sources
- OWASP: LLM06 Excessive Agency (Top 10 for LLM Applications)
- Cloud Security Alliance and Token Security: AI agent security incident research, 2026
- Gravitee: State of AI Agent Security 2026 Report
- Anthropic: Disrupting the first reported AI-orchestrated cyber espionage campaign (Nov 2025)
Figures reflect industry reports published in 2026. Check the linked primary sources for methodology and updates.