BORACAY — Businesses developing artificial intelligence (AI) agents need to secure their access to company systems and set limits on their actions, technology executives said, warning that even a harmless instruction could lead to confidential information being exposed.
Technology executives at the media briefing during the VSTECS CXO Tech Summit said some companies are already developing AI agents to help manage their computer systems. As these tools take on tasks, companies face decisions about what they can access, when employees should intervene, and how to recover if they take unintended actions.
“Sometimes agents or AI agents are treated just like any application. They forget that these agents also have their own identities that need to be secured,” said Bambi Escalante, country manager, Fortinet in the Philippines.
Escalante warned that an agent without sufficient safeguards could carry out actions beyond what a user intended while trying to complete an assignment.
“So maybe the initial instruction was not malicious, but because there were not enough guardrails for the agent, it went ahead and did a lot of things just to achieve whatever goal it was given,” she said.
Her warning followed comments from another executive about enterprises developing their own agents. The speaker identified governance as a concern, including preventing agents from deleting information or performing unnecessary actions.
“A lot of enterprises are already building their own agents,” said Walter So, managing director of HPE in the Philippines. “They don’t want these agents to start deleting things or going out of the box and doing unnecessary things.”
These agents help run IT systems, raising the need for rules to prevent unwanted actions.
So noted that agentic wallets are being developed in other countries that can make transactions on users’ behalf. He said businesses need to prepare the data and systems supporting these applications, with technology suited to their needs and size.
In other words, companies need to establish rules for agents operating within their systems, while keeping track of the AI tools employees use.
Escalante said that oversight should begin with visibility into approved and unapproved applications. Employee use of tools without company approval, known as shadow AI, can leave organizations unaware of where company information is being handled.
She said businesses should identify which applications are in use, who uses them, and who owns or manages them. Protection should also extend to the data and communications that agents rely on when working with other systems.
“So protect your data as well, protect your applications and your communication with other AI agents or other applications in the enterprise,” Escalante said.
Established security principles still apply, including zero trust, which requires checking access rather than automatically trusting a user or system, and defense in depth, which uses several layers of protection.
Companies also need records that allow them to examine AI activity. Escalante cited the ability to trace whether an employee uploaded confidential information, what actions occurred, and when they happened. Such records support incident investigations and reporting requirements.
Escalante said organizations should assess their ability to anticipate, withstand, and recover from an attack across their people, processes, technology, and governance.
“So there should be an assumption that a breach will happen,” she said.