Rubrik is implementing governance frameworks to manage rogue AI agents and prevent them from taking irreversible actions within enterprise environments [1, 2].
As organizations deploy autonomous AI to capture return on investment, the risk of agents performing catastrophic errors increases. Without strict oversight, an autonomous system could potentially delete a critical database or compromise system integrity, creating a tension between the need for innovation and the necessity of risk mitigation [1, 2].
Anneka Gupta, Rubrik's chief product officer, said these challenges were addressed in a June 2, 2026, discussion regarding the strengthening of governance to reduce risk [2]. The approach focuses on establishing guardrails that allow AI to operate effectively without granting the level of autonomy that could lead to permanent data loss [1, 2].
Effective governance requires a balance where the AI can still provide value and efficiency while remaining subject to human-centric or system-level checks. This prevents the "rogue" behavior that occurs when an agent interprets a command in a way that leads to destructive outcomes [1, 2].
Industry experience underscores the complexity of these deployments. Some experts in the field have worked in IT and cybersecurity for more than 20 years [3], noting that the transition to autonomous agents represents a significant shift in the threat landscape compared to traditional software automation.
Rubrik's strategy emphasizes that innovation does not have to be sacrificed for security. By implementing a layered governance model, companies can deploy AI agents that are capable of complex tasks, but are restricted from executing high-risk commands without explicit verification [1, 2].
“Rubrik is implementing governance frameworks to manage rogue AI agents”
The shift toward autonomous AI agents moves the primary risk from data breaches by external actors to operational errors by internal systems. By focusing on 'irreversible actions,' Rubrik is highlighting a new category of AI risk management where the goal is not just preventing bad data output, but preventing destructive system-level commands.



