5 ways AI will change how you defend your business in 2027

Google has published a security forecast for 2027 under its Cloud CISO Perspectives banner, arguing that agentic AI will reshape how businesses defend themselves within the next year. The company says that at the start of this year few organizations used large language models and AI agents in cybersecurity workflows, including vulnerability discovery and patching, and that this is now happening at scale. The forecast draws on commentary from Google Cloud executives, including Ramya Chitrakar, VP of engineering at Google Cloud Security, Heather Adkins, VP of security engineering, and Sandra Joyce, VP of Google Threat Intelligence.
The central claim is that agentic AI is outpacing the speed and scale at which traditional security operations centers can operate, and that this applies to attackers as well as defenders. The five trends Google lays out follow from that premise.
First, Google expects the alert queue to disappear. Alert fatigue is a familiar problem in SOCs, where analysts wade through large volumes of irrelevant or false positives. Google argues that autonomous, AI-driven threats can probe a network, gain entry and move laterally in minutes, which makes it impractical to chase alerts logged days or weeks earlier across multiple systems. Under the forecast, agents handle data enrichment, correlation and triage automatically, and the alert becomes the final output of the process rather than the starting ticket. Chitrakar is quoted as saying 2027 is the year analysts stop chasing alerts and start shaping outcomes. In practice, this reframes the analyst's job as oversight and verification of automated actions rather than first-pass triage, which is a meaningful shift in SOC staffing and tooling.
Second, Google says agentic AI needs identities and audit trails. Thousands of individual agents working across environments can become high-value entry points for attackers, so each agent should carry its own identity, much like an employee, and operate inside a zero-trust architecture. Adkins argues zero trust should extend past the network layer to the transaction layer. That means verifying each transaction and authorization, and giving every agent its own least-privilege credentials that can be revoked during an incident or when the agent is retired. Google's report notes that with real-time autonomous reasoning, a transaction allowed on Tuesday could be blocked on Wednesday if conditions look anomalous. For developers, this is essentially treating agents as non-human workload identities with scoped permissions, short-lived credentials and detailed logging, a pattern already familiar from service accounts and secrets management but harder to enforce when agents act autonomously across many systems.
Third, software development speeds escalate. Google contends security has to become a continuous system embedded in the delivery pipeline, because security work placed at the end of the development cycle and bug-fix tickets that take days no longer fit. Joyce notes that cybercriminals are poisoning everything from open-source code to software libraries, which AI coding assistants often pull automatically into codebases. The SolarWinds incident of 2020 is cited as a notable precedent for supply chain attacks. Google also calls for linking live production environments back to source code, so defenders can trace an exposure to the exact line of code, validate whether it is genuinely exploitable, and deliver a pre-tested fix at machine speed without breaking production. The implicit trade-off: faster, AI-assisted coding widens the surface that automated security controls must cover, so speed gains only hold if the pipeline checks keep pace.
Fourth, red-teaming must cover agentic AI. Google warns that organizations rushing to appear AI-first may treat security as an afterthought, exposing new attack surface through software vulnerabilities, misconfiguration, bypasses and indirect prompt injection, a class of attack where malicious instructions are hidden in content an agent processes. With AI now available to attackers ranging from script kiddies to state-sponsored groups, the forecast says penetration tests and red team exercises should include agentic AI and audit autonomous workflows before they reach production, then combine findings with threat intelligence to harden defenses.
Fifth, Google points to agent sprawl and the need for coordinated agent teams. Single generalist agents given broad instructions are unlikely to keep up with AI-assisted attackers that can compromise endpoints, enter networks, move laterally and steal data faster than before. Instead, the forecast recommends specialized, orchestrated teams of agents with their own domain training, memory and dedicated tools, each able to act within set limits under a central coordinator that knows when to escalate a decision to a human. Google frames unchecked agent growth as the next iteration of shadow IT. A likely trade-off is that orchestration adds management overhead and new failure modes, since the coordinator itself becomes a critical component to secure and monitor.
Why it matters: Security and platform teams should expect agent identities, least-privilege credentials, transaction-level verification and agent-aware red teaming to become standard requirements rather than optional hardening. For product teams building or embedding agents, the forecast suggests identity, logging and revocability need to be designed in from the start. Whether the 2027 timeline holds is an open question, but the direction of travel, toward treating agents as managed assets and risks, is one CISOs can plan against now.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Security leaders reveal the future consequences of AI outpacing human defenders.