Google Cloud warns AI agents will reshape cyber defence

Google Cloud has released its Cloud CISO Perspectives forecast for 2027, warning that autonomous AI agents will reshape cyber attack and defence.

Written by Chris Betz, Chief Information Security Officer at Google Cloud, the forecast sets out how agentic AI is changing the way security teams and threat actors operate.

AI agents are already being used at scale to find and fix vulnerabilities, cutting software release cycles from weeks or months to hours or days, Google Cloud said. It argued this shift will change both the pace and focus of cyber security work.

One of the forecast's central warnings is that attackers may move away from direct data theft towards what it describes as decision manipulation. In this scenario, threat actors would poison data or compromise business logic so autonomous systems make harmful choices, including on contracts, inventory routing or capital allocation.

The report also pointed to software supply chain risks tied to agentic development tools. It said attackers are targeting open-source software, external libraries and MCP tools that coding assistants can automatically pull into corporate codebases.

Patch cycle

Another theme is the decline of traditional patch management cycles. Multi-stage attacks now unfold in seconds, Google Cloud said, leaving weekly vulnerability scans and month-long patching processes too slow to respond.

Instead, it expects security controls to move further into software runtimes, with automated defences operating at machine speed. The forecast linked that change to a broader shift away from perimeter-based security towards systems that apply checks at the transaction layer across models, permissions, infrastructure and supply chains.

The report said this approach could ease the burden on human security teams by shifting them from alert handling to risk management. At the same time, it argued organisations will need tighter visibility into how autonomous systems act and why.

Human delay

Google Cloud warned that machine-speed attacks are erasing the window in which people can intervene. It said adversaries are exploiting vulnerabilities before vendors are aware of them and are using multi-agent frameworks to carry out large-scale credential harvesting in less than six hours.

According to the forecast, that pressure means defenders will need to rely more heavily on automated systems and structured governance. It argued that every AI agent should be treated as a first-class identity, with verifiable credentials and least-privilege access.

Google Cloud also said organisations should avoid relying on a single general-purpose agent for complex security tasks. Instead, it recommended teams of specialised agents, including hunter, decoder and responder agents, working with graduated autonomy and a central coordinator that can escalate issues to a human when needed.

Agent sprawl

A further concern is what the forecast called agent sprawl, which it described as a new form of shadow IT. The report said many enterprises are rushing to deploy agentic systems without putting in place the controls needed to govern and defend them.

For that reason, Google Cloud said companies will need full observability over agent decisions. That includes the ability to reconstruct inputs, reasoning, tool calls and policy compliance so autonomous systems can be validated and debugged.

It also said adversarial red-teaming should be mandatory before agentic systems are deployed. The forecast maintained that without complete traceability, organisations would struggle to trust the actions taken by those systems or identify where failures occurred.

Google Cloud said defenders still hold an advantage because they have access to internal context that attackers do not. The report cited identity privilege graphs, network architecture, code lineage and runtime data as sources that can help security teams identify dangerous permission combinations before they are abused.

The forecast was developed with contributions from Google Cloud and Google security leaders including Karthik Narain, Ramya Chitrakar, Heather Adkins, Sandra Joyce and Charles Carmakal.

"In 2027, there is a high risk that threat actors will shift from direct data theft toward decision manipulation, subtly poisoning data or compromising logic to prompt autonomous agents into making harmful business choices at scale," Betz said.