- The AI Trust Letter
- Posts
- Gemini 3: Is it the best model?
Gemini 3: Is it the best model?
Top AI and Cybersecurity news you should check out today
Rodrigo Fernandez
November 24, 2025
Welcome Back to The AI Trust Letter
Once a week, we distill the most critical AI & cybersecurity stories for builders, strategists, and researchers. Let’s dive in!
🤖 Google Unveils Gemini 3: A New Push Toward Agent-Ready AI
The Story:
Google announced Gemini 3, its newest generation of large models. The release focuses on stronger reasoning, faster performance, and better handling of long, complex tasks. The company says the model is designed to support the next wave of agentic systems.
The details:
Gemini 3 introduces major upgrades in long-context reasoning, allowing the model to work across large documents and extended workflows without losing coherence.
The model improves memory capabilities, helping it stay consistent through multi-step tasks and follow-up instructions.
Google says Gemini 3 is optimized for “agent behaviors,” meaning tasks like automation, planning, coding, and tool use should be more reliable.
The company is rolling out new safety evaluations and policies alongside the release, including stronger controls for sensitive queries and expanded red-teaming coverage.
Gemini 3 will be available across Google products, including Workspace, Android, Chrome, and the Gemini API, with wider availability planned for later in the year.
Why it matters:
Gemini 3 signals Google’s move toward AI systems that can operate as agents, not just text generators. Improvements in context handling, planning, and memory matter for any team building assistants, copilots, or automated workflows.
The parallel focus on expanded safety reviews suggests that Google expects these systems to take on higher-impact tasks and wants to demonstrate new baselines for responsible deployment.
⛓️💥 Cloudflare Global Outage: What Happened?
The Story:
Cloudflare published a detailed postmortem on the major outage that disrupted large parts of the Internet on November 18. The incident brought down multiple services for several hours and affected customers worldwide.
The details:
The outage was triggered by a software release to Cloudflare’s global control plane. A logic error in the update caused configuration data to propagate incorrectly across regions.
As invalid configuration spread, core networking and security services stopped functioning, including traffic routing and API operations.
Cloudflare initiated a global rollback, but the scale of the network slowed recovery. Some regions restored service quickly, while others took hours due to backlog and synchronization delays.
The company has paused similar releases, introduced new safeguards to prevent cross-region propagation errors, and is adding additional validation layers before future rollouts.
Why it matters:
Cloudflare sits in the middle of the modern Internet. A single faulty update can cascade into global downtime with real impact on businesses and users. The incident underscores how fragile large distributed systems can be, and why deployment controls, staged rollouts, and rigorous testing matter even for the most mature infrastructure providers.
🚨 Salesforce Customer Data Accessed After Gainsight Breach
The Story:
Salesforce notified customers that some of their data was accessed after a breach at Gainsight, a third-party platform used for account management and analytics.
The details:
The incident originated at Gainsight, where attackers used stolen credentials to access systems containing Salesforce customer data.
Exposed information varies by customer but may include names, email addresses, usage metrics, and other non-sensitive business details.
Salesforce says no internal systems were breached. The issue was limited to data shared with Gainsight for operational purposes.
Gainsight has reset credentials, added additional monitoring, and is conducting a forensic review. Salesforce is working with affected customers to verify what was accessed.
Why it matters:
This is another example of how breaches often hit through vendors rather than core systems. Companies relying on SaaS tools need visibility into which partners store their data, how those partners authenticate access, and whether third-party activity is monitored with the same rigor as internal systems.
👀 Cisco Flags Rising Risks From Aging Internet Infrastructure
The Story:
Cisco is warning that large portions of the global internet backbone are running on outdated hardware, making outages and performance failures more likely as traffic demands continue to grow.
The details:
Many core routers and networking systems in service providers and enterprises are reaching or past end-of-life, yet remain in production because replacements are costly and disruptive.
Traffic and compute loads have surged with cloud adoption, video streaming, and now AI workloads. Older equipment was never designed for these demands.
Cisco notes that maintenance windows are harder to schedule, meaning organizations defer upgrades and run equipment in “unsafe” conditions.
Aging components increase the probability of cascading failures, where a single device going down can spread across networks and impact multiple regions or services.
Why it matters:
Critical infrastructure failures don’t always stem from attacks. Sometimes the primary risk is simply equipment that’s too old to handle modern traffic patterns. As AI accelerates load on global networks, operators will need long-term plans to replace outdated hardware before it becomes a systemic weak point.
🔒 Building Safer AI: A must-attend Online Event
Learn how NeuralTrust helps enterprises deploy AI safely, combining AI runtime security, AI threat detection, and AI agent security. Also, discover real-world use cases, see how this technology is shaping the future of cybersecurity, and bring your questions. Join on Thursday, the 27th of November!
What´s next?
Thanks for reading! If this brought you value, share it with a colleague or post it to your feed. For more curated insight into the world of AI and security, stay connected.
