As extra firms deploy AI agents to hold out complicated duties inside inner techniques and transfer towards automation, some safety researchers have warned that these agents increase the assault floor from easy prompts to full enterprise techniques, posing a severe menace to organisations. This raises the query of whether or not AI agents will grow to be the brand new “insider menace.”
“AI agents are not inherently malicious. They are designed, constructed, and put in place to assist firms, not to trigger hurt. Nonetheless, they are relentless. These agents are extremely efficient at reaching their aims and discovering methods to perform the outcomes they are assigned,” John Scimone, President and Chief Security Officer, Dell, instructed indianexpress.com in an interview on the Dell Applied sciences World convention at the moment underway in Las Vegas.
“By means of that persistent and relentless deal with reaching these objectives, unintended penalties and behaviours can emerge if agentic implementations are not correctly ruled and safeguarded,” he added.
AI agents can browse the net, learn and write recordsdata, name exterior APIs, and question databases. Nonetheless, giant language fashions (LLMs) undergo from an unresolved flaw: immediate injection. Since LLMs do not clearly separate knowledge from directions, that means that any knowledge, such because the content material of an internet web page, an electronic mail, or a log entry, can successfully be interpreted as directions. The danger that follows is that attackers can manipulate AI agents from inside trusted techniques utilizing the agent’s personal id.
Gartner says that fewer than 5 per cent of enterprise functions used task-specific AI agents in 2025. In 2026, that quantity is anticipated to extend by 800 per cent. The analyst agency additionally estimates that greater than 40 per cent of enterprise functions will use AI agents in 2026.
John Scimone, President and Chief Security Officer at Dell, speaks in regards to the dangers posed by autonomous AI agents throughout Dell Applied sciences World in Las Vegas.(Picture credit score: Anuj Bhatia/ The Indian Specific)
“What we’ve seen internally and with our clients, and with so many organisations we’re partnering with throughout the globe that are actually leaning into this area, is that they are saying: we imagine on this know-how, and we imagine in the advantages it can convey. If you undertake a secure-by-design, secure-by-default method and put safety finest practices in place, you can do that securely and resiliently. In truth, you can obtain extra safety and resiliency than you may with legacy architectures and conventional infrastructure implementations,” Scimone mentioned.
Rising inner menace
As AI agents are given better entry to company networks and extra duties are allotted to them, the issue arises when they run 24/7 inside the community and, with their expanded capabilities, grow to be topic to dangers and threats.
Story continues beneath this advert
Wendi Whitmore, Chief Security Intelligence Officer, Palo Alto Networks, identifies autonomous AI agents as a significant rising inner menace.
In 2025, safety researchers confirmed {that a} prompt-injection assault uncovered Salesforce’s CRM platform to potential knowledge theft. Salesforce rapidly launched patches to stop AI agents from retrieving CRM data and sending them to exterior attackers. This vulnerability, dubbed “ForcedLeak,” illustrates how AI agents with out human oversight can be abused.
In one other case in 2025, safety researchers discovered a vulnerability in OpenAI’s Codex CLI coding agent that might permit attackers to execute malicious instructions on a developer’s machine by embedding dangerous directions in shared venture recordsdata. This might result in native system compromise, credential theft, code tampering, and potential downstream enterprise breaches, successfully turning the AI assistant into an assault entry level.
Security consultants warn that AI agents with broad system entry may increase the enterprise assault floor.(Picture credit score: Anuj Bhatia/ The Indian Specific)
Not solely do conventional safety measures now not work within the age of agentic AI, but a doable answer may be to deal with AI agents like people. You will need to perceive that when a human logs right into a system, they endure id verification, their actions are logged and monitored, and anomalous behaviour triggers alerts. Equally, AI agents also needs to be ruled with equal controls: a verified agent id, full logging of their actions and selections, behavioural monitoring, and full session monitoring.
Story continues beneath this advert
“As we go throughout that continuum of trade-offs between threat and reward with agentic AI, what we see is that each firm goes to have a distinct proper reply. However the frequent proper reply is having a course of by which you govern thoughtfully. You tie it to your online business technique, and also you tie it to your regulatory compliance necessities,” Scimone mentioned.
“It’s actually essential that as we set up the rules for threat administration, the rules for the governance that we are attempting to attain, that we’ve a method to really technically instantiate it. It’s not nearly coverage, but the precise know-how architectures are established to implement it and make sure that the agents behave as we design them to behave,” he added.
Source link
#agents #inherently #malicious #relentless #Dell #Chief #Security #Officer #John #Scimone
