Security for Mythos-Era Agentic Risks — with Rubrik’s Anneka Gupta and Cal Al-Dhubaib
Rubrik’s Anneka Gupta and Cal Al-Dhubaib discuss the escalating cybersecurity risks posed by powerful AI agents, the necessity of a "zero trust" mindset, and how platforms like Rubrik provide cyber-resilience through advanced visibility, governance, and remediation tools in the new agentic era.
The emergence of highly capable AI models, such as Anthropic’s Claude Mythos, marks a significant inflection point for cybersecurity. These models can find and exploit vulnerabilities in codebases at machine speed, a capability that is expected to proliferate to open-weight models within 6 to 18 months. This new reality demands a fundamental shift in how organizations approach security, moving from a focus on prevention and detection to an assumption of breach and a strategy centered on cyber-resilience.
The Dual Threat of AI Agents
The risk from AI is twofold. First, malicious actors can leverage these powerful tools to find and exploit vulnerabilities much faster than before. They don’t operate under AI usage policies, giving them a significant advantage. The second, and perhaps more insidious, threat comes from internal, well-intentioned AI agents. In their pursuit of completing tasks efficiently, these agents can inadvertently create security risks.
“AI agents, they’re very outcome focused. They’re going and trying to say, how do I get this job done in the least steps possible, the fastest way possible? and they’re not taking into account the same rules that us as humans, as employees working in an organization know to follow.” - Anneka Gupta
This creates a massive new attack surface. The proliferation of “non-human identities”—agents acting on behalf of users with access to keys, tokens, and APIs—represents an explosion in potential security gaps that traditional cyber controls were not designed to handle.
Embracing a “Zero Trust” World
The necessary response to this landscape is the principle of Zero Trust, which is essentially a framework built on the assumption of breach. It means architecting systems where every device, identity, and account is treated as if it has already been compromised. The goal is to ensure that even if an attacker gains a foothold, their ability to access critical data and infrastructure is severely limited. In practice, this involves minimizing access, making permissions ephemeral, and having a robust plan for what to do when—not if—a breach occurs.
Rubrik’s Cyber Resilience Platform
Rubrik’s platform is built to provide this resilience. It addresses the core questions leaders ask during a cyber attack: What was the impact? What was the blast radius? When did the attack start? How can we recover without reintroducing the same vulnerabilities?
The platform’s approach is built on three pillars, delivered through the Rubrik Agent Cloud:
- Visibility: The platform discovers and provides a holistic view of all agents operating in an environment (e.g., Microsoft Copilot, AWS Bedrock, Claude), what they are doing, and what permissions they have. This helps organizations understand their “posture up front” and identify potential risks, such as an agent having read/write access to multiple production databases.
- Governance & Control: This is managed by Sage (Semantic AI Governance Engine), a custom-trained small language model (SLM). Sage allows organizations to define policies in natural language (e.g., “agents should not give financial advice” or “agents cannot send emails to customers”) and enforces them at runtime. Rubrik chose an SLM over a frontier model to optimize for accuracy, cost, and low latency, which is critical for blocking actions in real-time without disrupting workflows.
- Remediation & Recovery: When an agent inevitably makes a mistake or a malicious action gets through, Agent Rewind provides the ability to undo those actions. Leveraging Rubrik’s deep expertise in cyber recovery and its “preemptive recovery engine,” this functionality can restore systems to their pre-incident state. It can rewind changes to production databases, fix broken relationships in a CRM like Salesforce, or revert unauthorized changes to identity systems. While it can’t recall a sent email, it provides a crucial safety net for system-level changes.
Trust Engineering: A New Discipline
Cal Al-Dhubaib introduces the Trust Engineering Framework, a discipline that combines human-centered design with robust trust infrastructure. It acknowledges that you cannot control models with 100% certainty and therefore must design systems to manage risk from the outset.
“Trust engineering is building off of what Anukica so eloquently described as the zero trust mindset but applied to AI systems.” - Cal Al-Dhubaib
This framework involves:
- Managing user expectations through clear signals (e.g., a self-driving car visually indicating it sees a pedestrian).
- Building trust infrastructure to monitor systems, detect misbehavior, and define fail-safes.
- Integrating people, process, and technology to ensure risk management is not an afterthought but a core part of the development lifecycle.
The rise of AI is blurring the lines between roles. Problems now cross the domains of the CIO, CTO, CISO, and Chief AI Officer. This necessitates cross-functional teams with shared literacy in AI risk, including roles like dedicated AI attorneys and data scientists versed in risk management. The future of AI-enabled knowledge work is, by default, cross-functional.