Google Cloud predicts agentic AI will accelerate cyber threats and defenses by 2027

man in blue dress shirt sitting on rolling chair inside room with monitors — Austin Distel / Unsplash

Google Cloud forecasts that autonomous artificial intelligence agents will significantly increase the speed of both cyberattacks and defensive measures by 2027. The company expects these systems to handle complex business functions such as pricing, vendor negotiations, and supply chain coordination. This shift requires organizations to verify the specific data, reasoning, and permissions used by agents to perform their tasks. The focus of corporate protection will expand beyond traditional data breaches and system outages to include the auditability of automated decision-making processes.

Chris Betz, the Chief Information Security Officer at Google Cloud, stated that vulnerabilities related to large language models and agent-based detection are spreading from early adopters to the entire enterprise. Software deployment cycles are shortening from weeks or months to hours or days. This compression creates a need to run development and security verification simultaneously. The company suggests that organizations must record the input data, reasoning basis, tools called, values passed, and policies that allowed specific agent actions. These records enable the reconstruction and verification of automated decisions, which is essential for scaling agent operations.

Heather Adkins, a Vice President of Security Engineering at Google, emphasized the need to extend zero trust principles to individual transactions and requests. Sharing human accounts or administrator tokens among multiple agents allows a compromise in one agent to spread to other systems. Each agent requires a verifiable identity and the minimum permissions necessary for its specific tasks. Systems must be able to block requests if unusual signs appear at the time of execution, even if the request was previously valid.

In security operations centers, AI agents are expected to collect logs from multiple systems and analyze correlations between alerts to provide initial investigation results. Human operators will define the scope of automated responses and the conditions requiring human approval. Google Cloud proposes an operational model where operators can understand the basis for an automated response within 60 seconds and revert the action within 30 seconds if necessary.

The company also identified risks in the development phase, where AI coding tools automatically fetch open-source libraries and tools based on the Model Context Protocol. These components can become targets for supply chain attacks. Google recommends tracking risks found in production environments back to the source code and verifying the actual impact and potential service disruption before deployment. Analyzing permission structures, network configurations, code change history, and real-time activity records helps identify risks that might be missed when examining individual settings alone.

As endpoint security improves against file-encrypting ransomware, criminal groups are expected to focus on quickly exfiltrating data and demanding money for its non-disclosure. Enterprise-developed AI model weights, agent capabilities, and prompt libraries are potential targets for theft. The unauthorized use of cloud GPU computing resources, referred to as LLMjacking, is also cited as a threat. Nation-state espionage organizations may avoid high-speed automation to reduce detection risks during long-term operations. Charles Carmakal, the Chief Technology Officer at Mandiant Consulting, stressed the importance of red team exercises to verify safety mechanisms and permission controls before deploying autonomous agents into real business workflows. Google warned that decentralized agent adoption can create shadow IT, making it difficult for central teams to maintain oversight. Organizations must clearly define the roles of detection, analysis, and response agents and determine at which stages human judgment is required before deployment.