The rapid proliferation of AI agents within corporate environments has created a significant visibility gap, as employees increasingly deploy automation tools without oversight from security or IT departments. Recent industry data indicates a disconnect between perceived and actual control: while a majority of security professionals believe they have strong visibility into AI agents, a much larger percentage report discovering unauthorized agents within their networks over the past year. These agents, often built using no-code platforms and frontier models, frequently operate with broad access to sensitive enterprise systems, including email, ERP software, and communication platforms like Slack.
The security risks associated with these agents stem from how they are granted access to internal data. Practitioners have identified four primary access methods: riding a human’s existing session via browser extensions, utilizing delegated OAuth grants, employing static credentials such as API keys or service account passwords, and using purpose-built, short-lived agent identities. The latter is considered the most secure but remains the rarest. A common danger is the use of OAuth grants or static keys that lack expiration dates, allowing agents to retain access to sensitive systems long after the original creator has changed roles or left the company.
A significant vulnerability arises when agents are designed to process untrusted input, access private data, and execute actions without human intervention. Security experts advocate for the Agents Rule of Two, a framework suggesting that an agent should never combine more than two of these capabilities in a single session. When an agent possesses all three—such as one that reads external emails, accesses vendor bank details, and has the authority to modify those details—it becomes a high-risk vector for malicious activity.
The threat of prompt injection is particularly acute because it can occur through content that no human ever reads. In one scenario, an agent tasked with summarizing emails was compromised when it processed an invoice containing hidden, malicious instructions embedded in white text. Because the agent was authorized to interact with the ERP system, it executed the fraudulent request to update vendor bank information. Because the agent used a personal API key belonging to its creator, the logs appeared to show legitimate, authorized activity, making the unauthorized change difficult to detect through standard permission-based monitoring.
Detecting such incidents requires moving beyond traditional permission-based baselines. Experts recommend baselining agents based on their actual behavioral patterns, such as the specific tools they call and the sequence of their operations. A compromised or rogue agent can often be identified by deviations from its established routine, such as a sudden transition from read-only tasks to write operations, or an unusual sequence of events—like reading an email followed immediately by modifying a payment record.
When a security incident involving an agent is suspected, responders are advised to suspend the agent rather than destroy it. This allows for the preservation of transcripts and memory, which are critical for forensic analysis. The challenge of post-incident auditing is significant, as many organizations lack the necessary records to answer fundamental questions: which agent performed an action, under what entitlement, with whose approval, and on whose authority. Without a clear audit trail that links the agent’s actions to a specific human requester and a verified delegation chain, organizations struggle to reconstruct the context of an event after the fact.





