A significant disconnect exists between confidence and reality among IT and security leaders regarding AI agents. Most believe their teams can identify when an AI agent goes rogue, yet swift mitigation remains a challenge. A recent survey from WanAware reveals that while 90% of IT leaders express faith in their capability to spot malfunctioning agents, only 26% can gauge the downstream effects within minutes. Over 45% report it might take hours to fully grasp the consequences of such incidents.
This timing gap is concerning. Jeffrey Collins, CEO of WanAware, indicates that the ability to respond quickly is critical, as issues stemming from malfunctioning agents can escalate rapidly, leading to outages or data breaches within seconds. Collins emphasizes that recognizing an incident is only half the battle; the real problem arises when understanding the full scope takes days or even longer. “If your average time to just knowing about an event is measured in days, weeks, or months, you have a serious problem right now,” he states.
Speed of AI Agents
Kevin Paige, field CISO at C1, reinforces the urgency in addressing problems with AI agents, stressing that these systems operate at machine speed. “The gap between an agent malfunctioning and you catching it isn’t measured in minutes; it’s measured in actions,” he explains. Because these agents often exploit previously granted permissions, they can propagate damage across systems before anyone realizes an issue has occurred, often revealed through external sources instead of internal detection tools.
This lack of proactive detection is problematic. Organizations frequently learn about rogue agents from customers or through issues that arise in downstream systems. Paige highlights the broader implications: “One incident like that and the business pulls back on agents entirely, so failing to contain a malfunction fast is also what stalls adoption.” For many, visibility exists, but practical control is often absent.
When an AI agent steps outside its designated boundaries, it typically does so without clear alarms. “Usually, it’s using access it legitimately has for a purpose nobody signed off on,” Paige notes, which eludes traditional access models. The aftermath often requires manual fixes, further complicating the response process. Chris Camacho, COO of Abstract Security, adds that solutions requiring granular control are essential: “Every agent should have its own identity, narrowly scoped permissions, and a complete audit trail.” Immediate control measures, such as revoking permissions during issues, are vital for effective management.
Overconfidence in Detection
The survey highlights a gap between expected and real capabilities in addressing AI agent malfunctions. Joe Brinkley, director of offensive security research at Cobalt, connects this overconfidence to the nature of compliance documentation rather than genuine operational readiness. “Tracing agent impact fast is brutal,” he explains. AI systems operate with nondeterministic logic across multiple APIs, making the complete execution chain difficult to trace through traditional logging. By the time alerts are triggered, agents may have already executed various downstream actions, complicating the incident response.
Vulnerabilities often manifest in the form of data flow issues. For example, a prompt injection can lead to unintended actions, turning a compliant AI into a hazard. “The agent suddenly thinks its official job is to dump your database,” Brinkley explains, highlighting the urgency for clear understanding of AI behaviors. Furthermore, issues like loop failures can exacerbate problems, where an agent repeatedly hitting a failing API drains resources and functionality.
To combat this, Brinkley advocates for implementing “hard kill” switches at the API layer to halt any agent that strays beyond its scope. He warns against relying solely on soft guardrails, pushing IT leaders to treat rogue agents like compromised user accounts, swiftly revoking access and pulling tokens to stop the threat immediately.
The core takeaway here reflects a need for organizations to reassess their AI management strategies. Those achieving success won’t merely deploy numerous agents but will ensure that they can articulate each agent's actions, confirm adherence to policy, and swiftly revoke control when necessary. As AI becomes increasingly integrated into business processes, striking the balance between confidence and operational capability will be crucial for sustaining trust and functionality in tech ecosystems.
