ZDNET’s key takeaways

  • Google expects agentic AI to influence security at scale in 2027.
  • Security experts say AI outpacing human-only response teams is our new reality.
  • Agents need an identity and must be managed as assets and cybersecurity risks.

At the beginning of the year, few organizations were utilizing large language models (LLMs) and AI agents in their cybersecurity workflows, including for discovering and patching vulnerabilities. According to Google, this is now happening at scale.

Also: Rogue AI incidents hit ‘tens of thousands’: Can businesses trust these tools?

The capabilities of agentic AI are fast outstripping the speed and scale at which traditional security operations centers (SOCs) can operate. Like it or not, AI has an integral role in the new reality — for the defenders and threat actors.

Speed and scale are the main reasons enterprises consider adopting AI in SOCs and for security more broadly, and these factors underpin Google’s new security forecast for 2027 — Cloud CISO Perspectives. However, without the right precautions, AI security adopters could inadvertently increase their attack surface rather than shrink it.

1. Threat alerts in your SOC move up the chain

Alert fatigue is well-known across SOC teams. Alert after alert needs to be trawled through; many are irrelevant or false positives, tying up skilled staff with the mundane, tedious, time-consuming task of triaging them.

Now that autonomous, AI-driven threats can probe a network, gain entry, and move laterally across it in minutes, chasing threat data and security alerts from multiple systems logged days or weeks ago can’t help a business mitigate them — and certainly not at the speed of the threat.

According to Google’s forecast, AI will mean the alert queue quietly disappears. Ramya Chitrakar, VP of engineering at Google Cloud Security, believes that “2027 is the year analysts stop chasing alerts and start shaping outcomes.”

Also: AI is both a cyber weapon and a massive target

How? CISOs must change the alert model itself. Instead of human defenders triaging alerts at the start of a cybersecurity incident, the alert becomes final output, with autonomous agents handling “data enrichment, correlation, and triage automatically, surfacing a verified conclusion rather than a raw ticket,” according to Google.

This means cybersecurity experts can oversee alert handling and verify automated actions, potentially freeing them to focus on other tasks.

2. Agentic AI needs identities and paper trails

Custom agentic AI is often designed — or at least intended — to work autonomously, with human oversight. However, having thousands of individual agents working across multiple environments can create high-value entry points into corporate networks that cyberattackers target.

We need to think of agentic AI as more than a tool: each one needs its own identity, like a human employee, to operate within a zero-trust architecture, retain control, and reduce the risk of compromise.

Also: Your enterprise AI access can be sold for pennies – beware ‘LLM Mixers’

However, Heather Adkins, VP of security engineering at Google, thinks this needs to go further and suggests the zero-trust concept shouldn’t stop at the network layer. Instead, Adkins believes it needs to reach the transaction layer.

A paper trail for each agent needs to be established, which is impossible to handle manually, and even within a zero-trust framework, AI agents could be given too much scope. Instead, implement verification for each transaction and for every authorization, and assign each AI agent its own least-privilege set of credentials that can be revoked if there is a security incident or the agent is no longer required.

“When backed by real-time autonomous reasoning, we can dynamically allow a transaction on Tuesday and block the exact same request on Wednesday if conditions look anomalous,” the report notes.

3. Software development speeds escalate

AI is being used to speed up the development and deployment of new software, and in 2027, security will have to become a continuous system embedded in the software delivery pipeline, according to Google.

Threat actors have carried out software supply chain attacks over the years, one of the most notable being the SolarWinds incident of 2020. Sandra Joyce, VP of Google Threat Intelligence, says cybercriminals are now poisoning everything from open-source code to software libraries, which AI coding assistants often pull automatically into codebases.

Also: How AI security failures are endangering your supply chain

Bug-fix tickets that take days to respond to and security appearing at the end of a software development cycle no longer work. Instead, writing and securing code must happen simultaneously, with the help of agentic AI, so speed doesn’t compromise security.

“It’s also crucial that we link live production environments directly back to source code so that defenders can trace a real-world 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,” Google added.

4. Red-teaming agentic AI

As organizations rush to implement AI into their applications and services to appear “AI-first,” innovative, and modern, security may be ignored or treated as an afterthought.

As a result, we may face a new, broader attack surface than we realize, as AI introduces new security issues, including software vulnerabilities, misconfiguration, bypasses, and indirect prompt injection attacks. Now that AI is also in the hands of all cyberattackers — from script kiddies to state-sponsored groups — auditing corporate networks and infrastructure is crucial.

Also: Who’s responsible for catching rogue AI agents? You are

Penetration testing and red team exercises now need to cover agentic AI. This wasn’t in scope for many red teams only a few years ago, but now that businesses have run with the idea of implementing AI in their workflows, processes, and products, organizations need to identify and fix security issues introduced by AI before threat actors take advantage of them.

CISOs should task pen testers with auditing before an autonomous workflow hits production, then combine the results with threat intelligence to ensure defenders have the information they need to bolster network defenses and harden classifiers against new and novel attacks.

5. Security at scale

The final significant trend we are likely to see in 2027 is tackling agent sprawl to create coordinated teams of AI agents that can rapidly handle security tasks at scale.

Single AI agents given general instructions won’t cut it when AI-assisted attackers can compromise endpoint devices, enter networks, move laterally, or start stealing data in a far shorter timeframe than before. Instead, CISOs should replace these “generalist” AI agents with specialized, orchestrated teams of AI agents that can meet the demands of defending against AI-driven attacks by employing specialist tooling and skill sets.

Also: AI agent kill switch urged by Okta-led alliance

CISOs also need to ensure that each agent has its own domain training and understanding, memory, and dedicated tools, and can act on its own within set limits, with a central coordinator that knows when to hand a decision to a human. This requires oversight, but in return, organizations will avoid agent sprawl, which the tech giant believes is the next iteration of shadow IT and its associated threats.