Hackers are increasingly using AI to do more than generate phishing emails or write malicious code. They're deploying autonomous, multi-agent systems that can conduct multiple stages of a cyberattack with limited human involvement.
Google's Threat Intelligence Group warned this week that bad actors’ growing use of these tools marks a turning point in how they carry out cyberattacks.
Among its findings, the group observed attackers using multi-agent penetration-testing frameworks during actual intrusions, including attacks targeting U.S. aerospace organizations. Researchers also identified a framework designed to analyze software patches, discover vulnerabilities, create exploits and execute them without any human involvement.
The researchers warned that the speed of these systems is reducing the time defenders have to respond.
For businesses, preventing an attack is only part of the problem. They also need to be able to stop an agent's activity, determine what it changed and restore affected systems.
The risks aren't limited to hackers. AI agents operating inside businesses can also make mistakes, including deleting or modifying data they were authorized to access.
Forrester research found that nearly a third of organizations have experienced a significant disruption related to an AI failure. Researchers also warned that enterprise agents could cause damage by deleting critical databases or compromising systems needed for recovery.
Amy DeMartine, vice president and senior research director at Forrester, addressed those concerns at the firm’s Security & Risk Forum earlier this month, warning that companies can't count on safeguards to catch every AI failure. Her advice was to make sure existing recovery plans account for AI-related incidents.
Giving an AI agent access to a system is one thing. For CIOs and CISOs, knowing whether the company can rebuild if that agent deletes critical data or makes changes that can't easily be undone is another.
Also in AI news this week:
Anthropic's IPO shows AI's value, but enterprises should stay wary: The generative AI vendor raised $81.3 billion in a record-setting IPO Thursday, becoming the first major model developer to go public. Its debut highlights investor enthusiasm for AI, although questions remain about enterprise returns and the company's reliance on major cloud providers.
Microsoft unveils Copilotforce, expands Office inside Copilot: At its Ignite conference, the vendor expanded Copilot's workplace capabilities, made Office inside Copilot generally available and introduced a Salesforce integration aimed at bringing AI into everyday business workflows.
Nvidia backs AI infrastructure buildout with $118B in guarantees: The GPU provider disclosed up to $118 billion in financial guarantees tied to AI infrastructure projects, taking on a larger role in helping customers secure the power and facilities needed to deploy its systems.
AI, rising costs top list of 2027 talent concerns, per Gartner: The firm’s researchers identified AI and rising costs as major workforce challenges heading into 2027, with particular concerns about the future of entry-level positions as companies rethink their staffing needs.
Boston Dynamics, GlobalFoundries expand robotics collaboration: GlobalFoundries plans to expand its use of Boston Dynamics' Spot robots across its semiconductor manufacturing operations and begin testing Atlas humanoid robots in early 2027.
Target's AI shopping assistant is driving sales: The retailer says shoppers using its Ask Target AI assistant are converting to purchases at roughly twice the rate of other shoppers, offering an early indication of how AI-powered shopping tools could influence sales.
Liz Hughes is an award-winning editor and writer covering AI and emerging technology and the former editor of AI Business and IoT World Today.