Earlier this year, an AI agent was given a simple task: clean up a codebase to improve efficiency. It deleted the entire thing. The agent wasn't malfunctioning. It was optimising. This story sits at the heart of a new whitepaper, “Power without control: Rethinking cybersecurity for the age of agentic AI”, produced by Economist Enterprise and supported by Rubrik. Its central argument is that the autonomy that makes AI agents transformative is the same quality that makes them dangerous. Organisations that treat cybersecurity as a matter of building higher walls will find those walls increasingly irrelevant. The imperative now is resilience—the capacity to absorb, recover from and adapt to failures that are no longer a question of if, but when.
The whitepaper was launched in New York this June at an invitation-only gathering titled ‘Governing Autonomy in the Agentic Era’, bringing together senior leaders across technology, security and AI to tackle one of the industry’s defining tensions: how to let AI systems act independently without losing control. Following a presentation of the research findings by Vaibhav Sahgal, principal and editorial strategy lead at Economist Enterprise, audience members listened to discussions with Kavitha Mariappan, chief transformation officer at Rubrik; Alan Rosa, chief information security officer at CVS Health; and Brian Saluzzo, chief information officer at Citi. Participants left with practical governance frameworks and resilience strategies they could put to work immediately.
Scaling agents, multiplying risks
The whitepaper, based on a global survey of over 800 business decision-makers, makes for sobering reading. Ninety-eight percent of organisations have already experienced a disruptive agent-related incident. Nine in ten expect more disruptions ahead, regardless of the safeguards currently in place.
Yet many organisations are operating under an illusion of preparedness. Ninety-five percent claim to have set recovery targets, but 42% acknowledge those targets are informal or loosely defined. More troubling still: three-quarters of organisations do not report cyber recovery performance metrics to senior leadership. Boards are making aggressive AI investments without visibility into whether the business could actually survive a serious breach.
As CVS Health’s Alan Rosa put it: "The conversation a cybersecurity team has with their board has shifted from 'I will prevent the breach' to 'there will be a breach'."
98%
of organisations have already experienced a disruptive agent-related incident.
9 in 10
expect more disruptions ahead, regardless of the safeguards currently in place.
Yet many organisations are operating under an illusion of preparedness. Ninety-five percent claim to have set recovery targets, but 42% acknowledge those targets are informal or loosely defined. More troubling still: three-quarters of organisations do not report cyber recovery performance metrics to senior leadership. Boards are making aggressive AI investments without visibility into whether the business could actually survive a serious breach.
As CVS Health’s Alan Rosa put it: "The conversation a cybersecurity team has with their board has shifted from 'I will prevent the breach' to 'there will be a breach'."
The conversation a cybersecurity team has with their board has shifted from 'I will prevent the breach' to 'there will be a breach'."
Alan Rosa, chief information security officer at CVS Health
The threat is now inside the perimeter
For years, cybersecurity strategy rested on the premise that a stronger perimeter meant a more secure organization. Agentic AI has quietly dismantled that logic. As Rubrik’s Kavitha Mariappan observed, AI agents have effectively unleashed a fleet of insider threats at an exponential scale, operating from within, at speed, with legitimate access.
The result is a crisis of confidence that is actively deepening. Recovery confidence across the industry is declining. The volume and variety of threat vectors now outpace traditional risk prioritisation. Panellists agreed that absolute prevention is no longer a credible goal.
Agents have effectively unleashed a fleet of insider threats at an exponential scale, operating from within, at speed, with legitimate access."
Kavitha Mariappan, chief transformation officer at Rubrik
Control at machine speed
AI agents operate without fatigue and act faster than any human oversight model can track. Panellists highlighted that keeping pace demands automation and risk-based prioritisation, but speed without precision creates its own vulnerabilities.
Citi’s Brian Saluzzo offered a critical insight: negative instructions do not work. Telling an agent "do not delete my code" is unreliable, as agents do not process prohibitions consistently. The answer lies in zero-trust controls applied directly to the underlying data, not to the agent's behaviour. Organisations also need a robust control plane that captures the full lineage of AI actions—who initiated the query, which models were involved, what actions were executed—so inappropriate operations can be intercepted before they cause harm.
Rollback and red teaming
Simulating attacks is paramount to surviving in this environment. Alan Rosa’s team at CVS Health runs numerous intensive, third-party tabletop exercises annually, alongside quarterly disaster recovery tests against critical systems, for example. Yet, testing alone is not enough; organisations desperately need robust rollback capabilities.
Without these provisions in place, you run the risk of ruining your reputation in one outage.”
Brian Saluzzo, chief information officer at Citi
Panellists stressed the absolute necessity of having immutable, air-gapped backups not just for data, but also for system configurations and identity ecosystems. Without an isolated recovery environment, threat actors or rogue agents will simply log back in and re-infect the system the moment it is restored. However, the research shows that, currently, seven in ten organisations lack the fully tested capabilities to reverse harmful agent actions reliably. And the risks of not having these in place are notable. As Brian Saluzzo pointed out, “without these provisions in place, you run the risk of ruining your reputation in one outage.”
Assuming breach
US$200bn
research estimates that the security industry recently spent in excess of US$200bn on cybersecurity
The clear consensus from the experts is that the pursuit of perfect prevention has become the industry's most expensive failure. As Kavitha Mariappan highlighted, research estimates that the security industry recently spent in excess of US$200bn on cybersecurity, including prevention and detection technologies to build taller perimeter walls. Yet the volume of enterprise incidents continues to grow faster than this spending. In the age of agentic AI, where autonomous systems operate continuously from within, fortifying those walls is no longer a viable substitute for fixing the structural foundations of a business. Security strategies must actively adopt an ‘assume breach’ mindset, which requires working backwards from an inevitable compromise to determine what is minimally viable to survive.
Ultimately, navigating the agentic era requires a fundamental reallocation of focus. While investment heavily favours prevention today, true resilience means accepting that failures will happen and shifting resources towards maintaining the ability to rapidly roll back, keep a clean state and continue operating.
