Most firms hit by AI agent security incidents, report

IRONSCALES has published a report finding that 77% of organisations experienced an AI agent-related security incident in the past year. The study was based on responses from 501 US IT and cybersecurity professionals.

The report found that 96% of organisations use AI agents or tools that can act on employees' behalf, while 57% reported at least one confirmed case in which an AI agent or tool acted on malicious, misleading or unauthorised content.

The findings point to a shift in phishing and social engineering threats as companies give AI systems access to internal tools, communications and financial workflows. Among organisations that had an AI-agent-related incident, the most common cases involved an AI tool summarising or acting on a malicious email or message, cited by 42% of respondents, and attackers using AI tools during phishing or social engineering attempts, cited by 41%.

Access levels appear broad. The report found that roughly four in five organisations give AI agents access to internal systems, 49% allow them into email inboxes or messaging platforms, and 30% grant access to finance, billing or payment systems.

Approval gap

The survey also found that security teams often play a limited role in approving AI agents, plugins and connectors. IT was named as the primary approver at 78% of organisations, while security held that role at only 6%.

Among organisations that experienced an AI-agent-related incident, 39% said an employee had given an AI tool access that security would not have approved. The report also found that only 52% of respondents said their organisation had formal policies, controls and incident response processes in place for AI agents, even though 93% said their incident response plan covers them.

That gap also showed up in confidence levels. While 84% said they were confident they could identify an AI-driven incident, more than three-quarters had already seen an AI agent act on harmful or unauthorised content.

Eyal Benishti, Chief Executive Officer of IRONSCALES, said the phishing threat model now includes software acting as an intermediary for workers.

"For years, the goal of a phishing email was to fool a person. Today it only has to fool the agent sitting between that person and their inbox," said Eyal Benishti, Chief Executive Officer of IRONSCALES.

"Organisations have handed AI agents the keys to their email, files, and even payment systems, and attackers are using the oldest tricks in the book to turn those agents against them. Social engineering now has two targets to exploit, people and agents, and modern security strategies must be able to reliably protect both," Benishti said.

Deepfake trend

The report also tracked changes in how organisations are responding to deepfake-related threats. The share investing in deepfake defence rose to 48% from 37% a year earlier.

At the same time, the proportion of financially affected organisations reporting losses of USD $100,000 or more fell to 49% from 63%, compared with the previous survey. IRONSCALES also said the estimated average loss among organisations that lost money to deepfake-related incidents declined from roughly USD $280,000 to roughly USD $225,000.

Seven-figure losses also fell, from 6% to 3%. Even so, deepfake incidents remained widespread, with 84% of organisations reporting some form of exposure, while the share describing deepfake defence as a top priority slipped to 60% from 71% as attention shifted to agentic AI.

The study suggests companies are dealing with two overlapping forms of cyber risk: AI-generated deception aimed at employees, and AI systems themselves acting on fraudulent material. More than half of respondents, 56%, said a deceived AI agent is at least as risky as a deceived employee.

Benishti said organisations should not let concern over one category of threat weaken defences against another.

"Last year's data showed what happens when organizations take a warning seriously: they invested, and deepfake losses became less severe," Benishti said.

"Agentic AI demands the same urgency, and it can't come at the expense of the progress already made against deepfakes. That starts with bringing security into the room when agents are approved and protecting the inbox for agents as well as humans," he said.