Dive Brief:
- Companies are struggling to oversee their AI agents because they don’t know what those agents’ normal behavior looks like, according to a new report.
- Nearly half (46%) of security professionals say they’re unable to establish safe baselines for agentic AI activity, the security firm Abnormal AI said in a report published on Thursday.
- The report highlights security gaps in AI oversight that businesses should prioritize as they increase their AI deployments.
Dive Insight:
Baselining is one of the most important steps any company can take to prevent AI agents from running rogue or being weaponized to sabotage operations or to steal data. In order for companies to refine the tools they use to detect potentially malicious behavior, experts say, they need to assemble a clear picture of what it looks like when agents are operating normally on their networks.
Companies understand this need. More than four in 10 businesses investing in behavioral-monitoring technology are directing it toward AI agents, according to Abnormal AI’s report — a far greater share than for any other corporate technology. Nearly one in three respondents said agentic AI was one of the most important areas for them to develop better baselining.
At the same time, more than half (53%) of respondents said their organizations didn’t establish normal baselines for machine identities — including AI agents — at all.
“It is the number one planned investment for practically every CISO I’ve talked to … for next year,” a top security executive at an unnamed B2B software firm told survey-takers. At the same time, the executive said, “there is no plan” for how to make that investment pay off. “We don’t know what the solutions are, but we’re all gonna put resources [toward] it.”
One of the difficulties of governing AI agents is that they blur the lines between human and machine behavior, because they sometimes take actions on behalf of human beings and sometimes act on their own initiative.
“That complicates a security model built around relatively stable relationships between identity, access, and action,” Abnormal AI said. Now, the security firm said, companies need to figure out not just whether an agent is allowed to do something, but whether it makes sense for an agent to be performing a particular task given its instructions.
Abnormal, which sells a behavioral-analysis platform, interviewed 525 security professionals in six countries for its report. Each respondent represented an organization with at least 1,000 employees. Eighty-five percent of respondents were CISOs or leaders of security operations centers.
Behavioral-analysis technology has limitations that undermine security professionals’ confidence in its findings, according to the report. Nearly four in 10 respondents said that a key obstacle to their ability to monitor network behavior was their difficulty in understanding or trusting the behavioral alerts they received.
The trust issue, the challenge of defining normal and the overwhelming number of false-positive alerts represent the three biggest hurdles that security professionals face when trying to analyze network behavior, according to Abnormal’s report. The security firm said it was notable that those three issues outranked more conventional obstacles such as staffing, budgets and tool capabilities.
