Google Cloud warns attackers will poison data to steer AI agents

Google Cloud released its Cloud CISO Perspectives forecast for 2027, written by its chief information security officer, Chris Betz. Google Cloud tells that AI agents, which are software that acts with no person, find and repair vulnerabilities at scale. It tells that this cuts release cycles from weeks or months to hours or days. It also tells that attackers use multi-agent tools to harvest credentials in less than six hours.

How it works

Betz warns of “decision manipulation”. Attackers poison data or change business logic, and agents then make dangerous decisions on contracts, inventory routing or capital allocation. The forecast also tells that attackers target open-source software, libraries and MCP tools that coding assistants pull into company code.

The proof

The forecast gives one time figure: credential harvest in less than six hours. It gives no count of attacks that poison data. Google Cloud tells that multi-stage attacks end in seconds, and weekly scans and month-long patch cycles are too slow.

Advice and gaps

Google Cloud advises these steps:

• Give each agent an identity with credentials and minimum access.
• Use teams of agents with one task each and a coordinator that sends hard conditions to a person.
• Record inputs, reasoning and tool calls, and do attack tests before launch.

The forecast gives no test results for these steps.

Insight: Vendors describe the speed problem that their products solve

Google Cloud tells that attacks at this time end in seconds, and a person cannot react in time. CrowdStrike launched security agents with no test results. A Gartner forecast shows 150,000 agents at one large company by 2028. This reads as a sales plan: vendors name machine speed as the problem, and sell agents as the answer. Buyers have more control if they get test results before they buy.