The rapid adoption of
generative AI and autonomous agents is putting AI developers and cybersecurity practitioners on a collision course — but the coming crack-up may end up benefiting both sides.
AI developers are trying to build increasingly capable systems that can reason, adapt and automate complex business processes. Cybersecurity practitioners are trying to make sure those same systems don't create additional unacceptable risk.
"People are rolling out AI agents at heavy rates," says Dev Rishi, General Manager of AI at Rubrik. "Those agents act without necessarily the full level of oversight that human employees have, but with the same hungry level of access and the same sort of non-deterministic behavior."
Now that organizations worldwide are embedding AI agents into everyday operations, these competing priorities are becoming very hard to separate.
Both sides are right. AI innovators need their agents to gain broad access to enterprise data and applications to unlock meaningful productivity gains. Security teams need confidence that those systems will behave safely and protect sensitive information.
Bridging these objectives calls for a unified approach to
AI governance — one that gives developers the freedom to innovate while providing cybersecurity professionals with the visibility, policy enforcement and runtime controls needed to manage intelligent systems responsibly.
"These two sides, they have the same end goal," Rishi says. "The way to be able to bring these people along is to have them to be able to both collaborate on a system that's allowing each to see how their actions are being used."
Where AI development and cybersecurity intersect
As organizations deploy AI agents throughout their businesses, security teams are discovering that AI workloads have become security workloads.
"AI is introducing a lot more new novel risk vectors that cybersecurity needs to be able to solve and patch," says Rishi. "AI is also giving cybersecurity professionals a whole host of new tools by which to be able to combat some of these types of risks."
Rishi notes that AI creates risks both externally, by accelerating
vulnerability discovery and exploitation, and internally, as organizations deploy autonomous AI agents with broad access to enterprise systems.
This convergence changes the role of cybersecurity. Instead of protecting applications, security teams must govern intelligent systems that can make decisions, access sensitive information and pursue objectives with limited human oversight.
How AI helps security professionals
Even as AI creates new risks, AI is becoming indispensable for mitigating those risks. Traditional security products rely on static rules and predetermined detections, but AI systems frequently behave in unexpected ways that conventional controls cannot anticipate.
"These models are quite creative, and it's very hard to be able to judge how they're going to operate," Rishi says, relating how Claude Code running on his laptop cheerfully leveraged Chrome to access Google Drive after he had blocked the AI's access.
He argues that organizations increasingly need AI to secure AI. Rather than expecting humans to review every AI action or relying on fixed rule sets, organizations can deploy specialized governance agents that continuously monitor AI workflows, detect risky behavior and intervene when necessary.
"I don't think you can rely on these static sets of rules or conventional software approaches to be able to secure and govern AI," Rishi says. "Instead, you need to be able to use AI to help you secure and govern AI itself."
When AI innovation becomes a security risk
Rishi points out a catch-22: AI development often becomes an information-security concern even before organizations can fully realize its business value.
"The unfortunate reality is that it becomes a security risk before the innovation happens in most organizations," he says. "Many Global-2000 enterprise organizations preemptively block some of the innovation because we don't have a way to be able to manage the risk."
Developers want AI agents connected to Salesforce, GitHub, financial systems and other high-value applications where the potential for large productivity gains is greatest.
"If we're looking for some of the productivity gains for AI, we need it to be able to display some part of what we are spending our manual time as human employees doing," says Rishi. "That means it needs to get the same level of access as those employees."
Security teams, meanwhile, worry that such practices will lead to unintended data exposure or AI misbehavior. Without effective governance, many organizations simply block these integrations, sacrificing potential business value in favor of safety.
The one area where there's no fear of AI is among tech startups, Rishi says, which understand that the rewards will likely far outweigh the risks.
"No Silicon Valley tech startup has any question about, 'Is AI delivering ROI?'" he says. "They're using it massively, and they're using it in a risk-forward posture because they're startups, and they're able to go ahead and plug it into all of these different systems."
Rishi also worries that as more companies block the use of AI for anything remotely risky, employees will simply find workarounds and use AI anyway. The term "shadow AI" is generally used for this sort of behavior under the assumption that employees are using unauthorized AI tools, but Rishi says that's misleading.
"I actually think the most risky shadow AI is the shadow AI that lives in the sanctioned tools that you're using," he says, mentioning that one client of his had five times as many MCP servers running as its IT team thought it had.
Why we need a unified framework for AI governance
Most cybersecurity controls were designed for deterministic software, not adaptive AI systems. Firewalls,
identity management and access controls remain essential, but they cannot fully govern AI agents that operate more like employees than traditional applications.
"AI … doesn't look much like static conventional software, where you put in one input, you get the same output," Rishi says. "It looks a lot more like securing and governing, almost in the same way you think about securing and governing humans."
So, as hinted above, Rishi says Rubrik is using AI to enforce those policies upon other AI agents with its Semantic AI Governance Engine, or SAGE.
"It's really hard to foresee every single thing that could go wrong with AI systems, and so you actually do need a more intelligent layer that's able to sit on top to be able to detect this," he explains. "It's essentially the guardian over the other types of agents that are being used."
An alternate proposal that's gained traction in the cybersecurity industry is to use existing identity-management systems to govern AI agents, which are a "third class" of identity combining human unpredictability with the speed of non-human identities.
Rishi, however, thinks it's a dead end to use old tools to try to manage AI.
"I actually don't think that the IAM infrastructure, the IAM security infrastructure, is even built to be able to secure and govern how agents interact," he says. "This is an example of why existing tools aren't good enough."
A unified AI governance framework, according to Rishi, would establish common definitions for acceptable AI behavior, runtime oversight and policy enforcement across clouds, endpoints and applications but frame them in ways that are very different from standard machine controls.
"The core applications that were built pre-AI aren't built to be able to support … granular policy understanding," he says. "The policies are much more similar to the types of policies you might imagine on a human enforcement level, rather than, 'Access control layer A says access to row B.' It's much more semantic."
Where AI innovators and security practitioners collide
"The fundamental conflict really boils down to one word, which is access," Rishi explains.
AI developers want agents connected to as many enterprise systems as possible because broader access enables greater automation and business value.
"We [developers] can't really guarantee something's not going to go wrong," he says. "We just want to be able to plug in the systems because we're very concerned about what could go right."
Security professionals, however, require assurances that granting this access will not expose sensitive data or create
unacceptable risk. Both groups share the same organizational goals, but they evaluate success through very different lenses.
"That's really the source of the underlying conflict. I want my agent to access all the different systems that I have. I want to be able to do all that work, and then I want to see what the agent can do," Rishi says. "On the security practitioner side, I do not want to give you that access unless you can give me guarantees."
The kind of controls that could bring together AI developers and security practitioners
Bridging this divide requires moving beyond blunt, all-or-nothing restrictions and toward intelligent runtime governance. Policies that evaluate context, intent and business risk as AI agents perform their work would give both developers and security practitioners a common operating model for responsible AI deployment.
"The way that the AI agent feels today, it's kind of like a fast car without brakes," Rishi says. "I don't know all the things that should be done to make it done securely. I want someone else that's an expert to be able to figure that out, and I want the ability to yell at them if they get it wrong."
He adds that developers would welcome a model in which security practitioners observe what they do and warn them if they're wandering off course.
"The right collaboration model is for security to be able to have a way to be able to define these policies and have a way that they can enforce these policies at runtime, and for the developer to feel like, 'What I want is I want to feel like something is watching.'" Rishi explains.
Under this model, security professionals establish governance policies, AI developers provide feedback when controls become overly restrictive, and specialized AI governance systems enforce those policies automatically. This lets developers innovate confidently while giving security teams visibility and assurance that AI agents remain within acceptable boundaries.
"Developers and security have never butted heads on what are reasonable restrictions," Rishi says. "The reason that they've actually butted heads is because security doesn't have a way to be able to enforce their reasonable restrictions, and so they have to do an overarching, overbearing restriction. That's what we're solving."
