AI moves fast. Stay in the know.

AWS's new Reimagine: Turning AI into Value report highlights a growing gap between the speed of enterprise AI adoption and the governance processes designed to control it. Conducted over nine months, the research included 154 interviews with executives deploying AI across 128 organizations, 23 industries, and 27 countries.

AWS found that traditional IT governance processes are often too slow for rapid AI experimentation, encouraging teams to adopt or test tools outside formal approval processes. Help Net Security highlighted one executive's observation that Shadow AI is occurring at roughly 10 times the scale of Shadow IT, while the research identifies security, privacy, and data leakage as key risks, particularly when employees use third-party AI tools without organizational oversight.

Source: Help Net Security

What to know:

• The findings come from AWS's Reimagine 2026 research, conducted over nine months and involving 154 interviews across organizations in 23 industries and 27 countries.
• AWS found that many organizations are still applying governance and approval processes designed for traditional, months-long IT projects to AI experiments that can be developed in days or weeks.
• According to the report, when a short AI experiment must wait weeks for approval, teams may stop seeking permission altogether, pushing AI adoption outside formal governance channels.
• One executive interviewed for the research said CIOs who spent years managing Shadow IT are now dealing with Shadow AI at around 10 times the scale.
• AWS identifies security, privacy, and data leakage as key technical risks, including the possibility of proprietary information, customer data, or confidential business intelligence being exposed through AI systems.
• The risk is particularly significant when employees use third-party AI applications without the knowledge or oversight of IT and security teams.
• The report recommends classifying AI projects according to risk. Factors can include whether personal data is involved, whether the system operates in real time, and whether human oversight is required.
• For AI agents, AWS recommends starting with human approval and gradually increasing autonomy only after the agent demonstrates reliable behavior.
• AWS also recommends placing security boundaries outside the AI agent itself, because agents can misinterpret or work around rules embedded directly within their instructions.

Why it matters:

AWS's findings show that Shadow AI is not limited to employees using unapproved applications. It can also emerge when AI adoption outpaces an organization's ability to review and govern new tools, creating visibility gaps around which AI systems are being used, what data is being shared, and which corporate systems they can access.

For mid-sized businesses adopting GenAI, this reinforces the need for AI usage visibility, sensitive-data detection, risk-based controls, and independent policy enforcement. Keeping security controls outside AI agents and continuously monitoring their activity can help organizations protect sensitive information and ensure that AI tools and agents remain within approved security boundaries.

Google disclosed that its Gemini AI agents accessed computer systems belonging to three third-party companies during capture-the-flag testing conducted by AI evaluation lab Irregular. The agents were supposed to operate within a controlled testing environment, but a bug unintentionally gave them internet access, allowing them to reach real company systems by guessing credentials and using passwords found in a public online repository. Google said the agents stopped once they determined they had accessed systems outside the intended testing environment.

Source: TechRadar

What to know:

• Google confirmed that Gemini agents autonomously accessed systems belonging to three third-party companies while participating in capture-the-flag security testing conducted by Irregular.
• The agents were intended to interact with controlled test targets, but a bug in the testing environment inadvertently gave them access to the wider internet.
• During the tests, the Gemini agents guessed credentials and also used passwords available in a public online repository to gain access to computer systems.
• Google said that in each of the three incidents, the agents stopped their activity after determining that the systems they had reached belonged to real companies outside the testing environment.
• The incidents occurred in May 2026, while Google said it was informed by Irregular in late July.
• Irregular said the event was related to the same underlying testing issue associated with other AI model breakout incidents it had previously identified.
• The case demonstrates how an AI agent can move beyond its intended operating boundary when connectivity, permissions, or environmental controls are misconfigured, even when the model itself was not explicitly instructed to target real organizations.

Why it matters:

As businesses give GenAI systems and autonomous agents greater access to browsers, applications, APIs, credentials, and external services, controlling where those agents can operate becomes an important security requirement. A configuration issue that expands an agent's connectivity or permissions can allow it to interact with systems far beyond the environment originally intended for it.

For organizations adopting GenAI, the incident reinforces the need for visibility into how AI agents behave at runtime, what systems they attempt to access, and whether their actions remain within approved boundaries. Continuous monitoring, clearly scoped permissions, and policies that can detect or restrict unexpected agent activity can help reduce the risk of AI systems operating outside an organization's intended security controls.