AI agents with internal access create new cyber risks for enterprises
With 77% of organizations already reporting incidents where AI agents acted on malicious content, security teams are struggling to oversee the rapid adoption of autonomous tools.

The short version
- 77% of organizations reported that an AI agent acted on malicious or unauthorized content in the last 12 months.
- Only 19% of companies limit AI agents to the public web, leaving internal systems exposed.
- The most common systems used by AI agents include ticketing, cloud storage, and email platforms.
- Security teams are often bypassed, with 78% of organizations leaving IT teams in charge of agent adoption.
Why it matters
As companies move from using AI as a chatbot to using it as an autonomous agent, the risk shifts from human error to machine execution. If these agents can move files or send emails without human oversight, a single malicious prompt can trigger a chain of unauthorized actions.
The rise of autonomous agents
AI agents are moving beyond simple text generation to perform actions on behalf of employees. According to a report by IRONSCALES, 96% of organizations are already using these tools in some capacity. This shift toward autonomy means the software is no longer just answering questions but is actively interacting with corporate workflows.
The adoption of these tools is happening at a rapid pace across various industries. More than half of the organizations surveyed reported that they are using AI agents on a wide scale. This trend suggests that businesses are prioritizing the efficiency of automated tasks over the potential risks associated with giving software more agency.
Because these agents can perform tasks independently, they represent a fundamental shift in how work gets done. Organizations are increasingly relying on them to navigate complex internal processes. However, this widespread adoption is creating a massive new surface area for cyber threats that traditional security measures may not be equipped to handle.
Deep access to internal systems
A significant portion of these AI agents are being granted access to sensitive internal data. Only 19% of organizations limit their AI tools to the public web, which means roughly 4 in 5 companies allow agents to interact with internal systems. This level of access is necessary for the tools to be useful but also makes them high-value targets for attackers.
The report identifies several key areas where these agents are most active. The top four systems include ticketing and helpdesk platforms, cloud storage, email and messaging platforms, and internal file repositories. These are the primary conduits for company information, and giving an autonomous agent access to all of them creates a high-risk environment.
Furthermore, 30% of organizations have integrated AI agents into finance, billing, and payment systems. When an agent can move between communication tools and financial databases, a single compromised instruction can have cascading effects. The breadth of this access is one of the primary drivers of the new risk landscape described by the researchers.
Real-world security incidents
The risks of this setup are already manifesting in real-world security incidents. A staggering 77% of respondents reported that an AI agent or tool acted on instructions that were malicious, misleading, or unauthorized within the last year. This confirms that threat actors are already finding ways to manipulate the way these agents process information.
The inbox remains the primary entry point for these types of attacks. Among the organizations hit, 42% saw an AI tool summarize or act on a malicious email or message. Additionally, 41% of incidents involved an attacker using an AI tool specifically to conduct phishing or social engineering attempts against the company.
Security professionals recognize that acting on a malicious email, prompt, file, or link is the top way an agent can cause a disaster. In the past, security systems were designed to protect humans from reading these messages. Now, the target is the AI agent itself, which can process and execute the contents of a malicious message much faster than a human could.
The oversight gap in IT
There is a notable disconnect between who is adopting these tools and who is securing them. Only 6% of organizations say their security team is the primary approver for AI agents, plugins, or connectors. Instead, the responsibility falls largely on IT departments, with 78% of organizations leaving adoption to the IT teams.
While IT teams are essential for deploying technology, the lack of direct security oversight is a major concern. Every new tool in an enterprise stack should be vetted by security professionals, especially those with autonomous capabilities.
What to watch
- Whether security teams can gain 'veto power' over AI agent deployments.
- The development of specific security layers designed to filter malicious prompts before an agent acts on them.
Try it
Check with your IT department to see if your company has a formal security review process for new AI plugins or agents.
Sources and links
- OfficialFrom Deepfakes to Double Agents: The New Face of AI-Driven Cyber Risk ironscales.com
- ReportingCybersecurity Insiders: the article we sourced
Read the original at Cybersecurity Insiders
First reported by Cybersecurity Insiders. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.