The rapid adoption of generative and agentic AI in cloud environments introduces significant security risks.
Bisham Kishnani, head of Security Systems Engineering for Asia Pacific & Japan at Check Point Software Technologies, explains how organisations can address critical security gaps created by rapid AI adoption in cloud environments.
Significant security risks
For Kishnani, the biggest risk is that attackers are now automating the entire kill chain with AI agents.
Anthropic’s research revealed that attackers have used AI coding agents to autonomously conduct reconnaissance, identify digital assets, scan for and exploit vulnerabilities, and exfiltrate data. In one case, a human operator provided a target name, and the agent executed the entire attack chain independently.
On the defender's side, this agentic pattern also poses internal risks: a compromised or manipulated agent can turn a single prompt injection into real actions such as querying data, calling APIs, or moving files. There is no way to restore from backup once a model or dataset is tainted or stolen.
“Addressing this requires runtime control, not just model-level safety,” Kishnani said.
The Check Point executive recommends addressing these risks with a unified layer that identifies where AI and agents operate, protects against prompt injection and data exfiltration in real time, and governs permissions and trust relationships across employees, applications, and agents. This approach treats an attack as a chain to be broken at any link, rather than a single point to patch.
Kishnani emphasises that these risks are not theoretical, citing DigiCert’s data showing that 78% of organisations have already reported AI-related incidents or identified an AI vulnerability in the past year.
Secure AI innovation
“The goal: builders move fast inside guardrails that are already switched on.” Bisham Kishnani
AI adoption is outpacing many organisations' governance frameworks. CISOs must enable AI innovation while ensuring security and risk controls keep pace.
“The gap is measurable,” admitted Kishnani.
Check Point's 2026 Cloud Security Report found that 77% of organisations have updated their cloud security strategy for AI, but only 26% have the architecture to enforce it.
“CISOs close this by making security part of the platform, not a review gate,” Kishnani shared.
He suggests enabling AI-aware inspection within existing infrastructure. Activating AI firewall capabilities at the network layer can discover agents, applications, and tools, inspect the intent behind prompts, and enforce DLP before a request reaches a model, all without requiring a new deployment project.
“It also means unifying oversight,” he added, noting that solutions can consolidate discovery, protection, and governance across employees, applications, and agents into a single view. This avoids splitting usage monitoring, application security, model review, identity, and data protection across multiple teams, each with only partial visibility.
“Governance also needs to run continuously and automatically — policy alignment checks, drift auditing, and exposure prioritisation at machine speed — because manual quarterly reviews can't keep pace with how fast new agents and models get deployed,” he added.
For Kishani, there is one goal: “builders move fast inside guardrails that are already switched on.”
Redesigning identity, access controls and data governance
AI systems increasingly require access to sensitive enterprise data. The next challenge for organisations is redesigning identity, access controls, and data governance to provide necessary access without increasing data exposure risk.
“Traditional IAM (Identity & Access Management) assumes a human logging in; it wasn't built for thousands of ephemeral agent sessions calling APIs on their own,” Kishnani said.
He suggests organisations need machine-native identity, assigning a unique, auditable identity to each agent, implementing least-privilege and just-in-time access for specific tasks, and using micro-segmentation to prevent compromised agents from exceeding their permissions.
“Like a badge that opens only the drawers a role needs,” he explained.
“This matters because every AI interaction is effectively a data-movement event,” Kishnani added.
Check Point Research's Annual AI Security Report found that organisations use approximately 10 AI applications per month, many of which are unapproved. High-risk prompts containing sensitive data have doubled year over year.
“Governance needs to inspect that movement in real time, not just data at rest,” Kishnani explained.
He noted that solutions exist to address both sanctioned and shadow AI use by applying live DLP to prompts and mapping which data and tools each agent can access.
“Layering identity across zones — users, applications, models, agents, and data/tools — each with distinct controls, means a breach in one zone doesn't cascade into full exposure of the next,” he explained.
"AI-ready" cloud security architecture
Organisations seeking a truly "AI-ready" cloud security architecture must understand how it differs from traditional security architectures.
“Traditional Zero Trust had three zones: users, applications, and data,” explains Kishnani.
He posits that AI-readiness involves adding two entirely new zones: AI orchestration/models and AI agents.
“Five new components now sit in the request path: the AI agent, the LLM itself, the RAG pipeline that fetches context, the GPU/AI factory infrastructure, and the MCP gateway connecting agents to outside tools,” he said.
Each is exposed to recurring risks in various forms, such as prompt attacks, supply-chain compromise, and agents acting improperly. Defences must therefore address the entire chain, not just a single link.
“AI-ready architecture applies defence-in-depth at three zoom levels (network, application, individual AI process) rather than a single static perimeter, with continuous, machine-speed enforcement instead of periodic patching,” he explained.
An AI Defence Plane that unifies Discover, Protect, and Govern into a single control plane can replace the fragmented setup where usage monitoring, application security, model testing, identity, and data protection are managed by separate teams without full visibility of the execution path.
“The practical shift: legacy security asked "is this connection allowed?" once; AI-ready security continuously asks whether this specific agent's specific action on this specific data is still appropriate right now,” he said.
Critical security decisions
The costliest mistake is underestimating how far attackers have already automated their side. Bisham Kishnani
Before deeply embedding AI into enterprise operations, CISOs need to make critical security decisions and avoid mistakes that could leave their organisations exposed.
“Three decisions matter most,” Kishnani said.
- First, decide how agent identities and the two new Zero Trust zones — AI orchestration/models and AI agents — will be governed before scaling deployment; retrofitting this after hundreds of agents are live is far harder.
- Second, decide where AI inspection sits in the network path — ideally inline at firewalls and API/MCP gateways already in place — so visibility doesn't depend on remembering to bolt it on later.
- Third, identify which data and models are true "crown jewels" and plan for irreversible loss, since AI offers no clean rollback once a model or dataset is exfiltrated.
“The costliest mistake is underestimating how far attackers have already automated their side: the disclosed case of an AI agent running an entire kill chain — recon, exploitation, exfiltration — from a single human instruction shows defenders can no longer rely on catching each stage manually,” he said.
Fragmented ownership across application, data, and infrastructure teams, along with the assumption that existing IAM covers agentic behaviour, increases exposure. He advised organisations to deploy a unified control layer to close this gap before attackers exploit it at scale.
