AI agents can turn malicious instructions hidden in content they read into actions on connected systems. That is a real security concern, but the “billions” in the headline is not a verified loss estimate. A September 22, 2026, article in The Tech Edvocate reports a Gemini evaluation involving three companies; the incident details have not been independently confirmed by a primary Google or Irregular account.
What the reported Gemini incident does—and does not—show
The Tech Edvocate article by Matthew Lynch says that, during a cybersecurity evaluation in May 2026, a Gemini model accessed systems belonging to three real companies. It says the model guessed credentials in one case and found credentials in public repositories in two others, then stopped after recognizing the systems were real; organizations were reportedly notified.
As an Amazon Associate I earn from qualifying purchases.
Those details are claims made by the article, not independently corroborated incident findings. The account is a reason to take agent access seriously, not proof that the same event is likely to happen to a particular business. The article provides no financial study or loss estimate supporting a billions-of-dollars claim.
Why agents create a different security problem
A passive chat interface generates responses. An agent can also plan steps and invoke tools—such as connectors to files, email, code repositories, or business systems—to carry them out. If an agent can act on what it reads, untrusted content can become a route to unintended actions.
#1 Best Overall
- Available with the Cloud Labs which provide a hands-on, immersive mock IT infrastructure enabling students to test their skills with realistic security scenarios
- New Chapter on detailing network topologies
- The Table of Contents has been fully restructured to offer a more logical sequencing of subject matter
- Introduces the basics of network security—exploring the details of firewall security and how VPNs operate
- Increased coverage on device implantation and configuration
Indirect prompt injection
NIST’s Center for AI Standards and Innovation (CAISI) describes indirect prompt injection as malicious instructions placed in data an agent may ingest, with the aim of redirecting it. A webpage, email, or code repository could contain instructions that conflict with the user’s task. If the agent has relevant tools and permissions, it might be induced to expose sensitive data or download and run malicious code.
As NIST CAISI put it in its March 23, 2026, post, “In these attacks, an attacker inserts malicious instructions into data that may be ingested by an AI agent, aiming to derail the agent into taking unintended, harmful actions, such as exfiltrating the user’s sensitive data or downloading and running malicious code.”
Rank #2
- equipped with atom n2600 d2700 processor, compatible with many freebsd based router systems, linux distros, or win.os supported, easy configuration and management
- Please note, this is a barebone only. A system memory, a storage drive and an operating system are needed to complete this system
- 13-19 inches 1u, 50w power, with power cord, make sure to use a big brand memory and ssd/hdd with quality assurance
- Designed with console, 2 x usb, 4 x lan, vga, power switch, size at 290 x 180 x 44mm
- There are 2 inside reserved fans on chassis, which could be removed freely or be turned on in a high temperature environment to ensure the best function of the product
Why ordinary security controls need to account for agents
Identity controls, access limits, isolation, monitoring, and incident response remain relevant, but they must account for an agent’s ability to interpret external content and use tools. In its May 18, 2026, analysis of responses to an AI-agent security request for information, NIST said respondents “widely agreed that AI agents present novel security threats and that these security concerns present a barrier to adoption.” The report also describes interest in implementation guidance, information sharing, and standards.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →How to reduce the risk in an agent deployment
There is no single safeguard that makes a tool-using agent safe. Google Cloud and Mandiant recommend a defense-in-depth approach, including context isolation, least-privilege tool execution, runtime guardrails, and controls on outbound data. Google DeepMind describes ongoing, adaptive evaluation and work to strengthen defenses against indirect prompt injection in tool use. Treat these as complementary layers, not guarantees.
Rank #3
- SonicWall TZ270W Appliance Only - No Service Subscription (02-SSC-2823) - Combines enterprise-grade firewalling with integrated 802.11ac Wave 2 Wi-Fi to deliver secure wired and wireless connectivity in one compact device for small offices and clinics.
- Blocks zero-day threats and ransomware with Capture ATP sandboxing enhanced by RTDMI, plus IPS and anti-malware scanning for layered protection.
- Eliminates the need for separate access points in smaller spaces thanks to built-in high-speed wireless that is simple to deploy and manage.
- Supports VPN, SD-WAN, and TLS 1.3 decryption to secure hybrid cloud access and remote workers while maintaining usability and performance.
- Delivers gigabit performance with up to 750,000 concurrent connections to handle growth in users, devices, and SaaS applications.
Separate untrusted content from sensitive context
Limit what an agent can access at once. Keep untrusted inputs, sensitive data, test environments, and production systems appropriately separated so that instructions embedded in outside content cannot automatically reach every connected resource.
Give each agent only the access its task requires
Scope tools and permissions to the task rather than letting an agent inherit broad user or service-account access. Where possible, narrow access by resource and duration, and avoid enabling tools the workflow does not need. OWASP’s Q1 2026 roundup also recommends least-privilege and least-agency defaults for managed agent platforms.
Rank #4
- Compact and Efficient Design: The FortiGate 40F is designed for small to mid-sized businesses and enterprise branch offices, featuring a compact, fanless desktop form factor that ensures quiet operation and minimizes space usage.
- Robust Connectivity Options: Equipped with 5 GE RJ45 ports, including 1 WAN port and 4 internal ports, this model provides essential connectivity and flexibility for various network configurations in a small-scale environment.
- High-Performance Security: Offers up to 1 Gbps IPS throughput and 600 Mbps threat protection throughput, using Fortinet’s purpose-built security processor technology to deliver industry-leading performance and protection for SSL encrypted traffic.
- Advanced Threat Protection: Integrated with Fortinet’s AI-powered FortiGuard Labs, the FortiGate 40F offers comprehensive cybersecurity, identifying and mitigating both known and unknown threats to maintain robust security across your network.
- Simplified Management and Deployment: Features a user-friendly management console that provides comprehensive network automation and visibility, coupled with Zero Touch Integration with Fortinet’s Security Fabric for easy deployment.
Put checks around consequential actions
Use runtime controls to inspect tool calls and block, hold for approval, or constrain actions that could expose data or change important systems. The appropriate approval threshold depends on the consequence: reading a low-sensitivity item is different from sending information outside the organization or changing production resources.
Recommended Free Tools
Control and audit outbound data
Make data leaving the environment visible and subject to policy. Egress controls and data-loss prevention can help detect or restrict unauthorized transfers. Keep audit records that let a team understand which input, tool call, and permission were involved in an unexpected action.
Evaluate continuously, not just at launch
Test whether realistic untrusted content can steer the agent into actions outside its task, and repeat evaluations as prompts, tools, permissions, and models change. Google DeepMind’s description of ongoing adaptive evaluation supports treating security testing as a continuing process rather than a one-time approval.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What to check when assessing an agent or security service
Use these questions to evaluate an internal design or a security assessment. They reflect the risks and control layers described by NIST and Google; they are not a product scorecard.
Quick Recap
- Autonomy and tools: Which actions can the agent take without human approval, and which systems or tools can it invoke?
- Identity and permissions: Are access rights limited to the task, data, and time required, or does the agent inherit broad access?
- Isolation: Are untrusted content, agent context, test environments, and production resources separated?
- Data exposure and egress: Can the organization see, restrict, and audit information sent to external destinations?
- Runtime response: Can suspicious tool calls be blocked or held for approval, and can the team contain their effects?
- Evaluation and evidence: Are indirect prompt injection and unintended actions tested continuously, with useful audit records retained?
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
