The First Leadership Compass on GenAI Defense

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!

🛡️ NeuralTrust Named Leader in KuppingerCole’s Generative AI Defense Compass

The Story:

KuppingerCole has released its 2025 Leadership Compass for Generative AI Defense, naming NeuralTrust a Market Leader, Innovation Leader, and Overall Leader. The report evaluates vendors helping organizations secure generative AI systems across runtime protection, red teaming, monitoring, and compliance.

The details:

• The report cites the rapid rise of AI misuse and the need for purpose-built defenses that go beyond traditional security tools.

• NeuralTrust is recognized for comprehensive coverage across the attack surface, including prompt injection defenses, agent security, data leakage prevention, and behavioral threat detection.

• The evaluation highlights our real-time AI Gateway and continuous monitoring capabilities, as well as our red teaming and evaluation solutions used to detect model weaknesses before deployment.

• KuppingerCole notes growing enterprise demand for unified AI security platforms instead of isolated controls.

Why it matters:

As organizations scale their use of generative AI, independent assessments are becoming essential for understanding which security controls actually work in practice. Recognition in this report validates the importance of treating AI systems as part of the core attack surface and adopting integrated defenses rather than piecemeal solutions.

Download for free

🤖 OpenAI Releases 2025 State of Enterprise AI Report

The Story:

OpenAI published new data on how enterprises are adopting generative AI, where value is emerging, and where organizations continue to face barriers in scaling deployment.

The details:

• Adoption has expanded across industries, with 92% of surveyed companies running at least one AI use case in production.

• The largest gains come from automation, customer support, and agent-style workflows that handle multi-step tasks.

• Governance remains a major bottleneck. Many teams report uncertainty around model oversight, data protection and compliance requirements.

• Skilled talent is another constraint, with organizations saying they lack enough engineers who can build, evaluate and maintain AI systems responsibly.

• Companies with formal evaluation and monitoring frameworks are moving faster and reporting higher ROI than those without them.

Why it matters:

The report shows that enterprise adoption is no longer the limiting factor. The real challenge is operational maturity. Companies that invest early in evaluation, monitoring and clear governance structures are better positioned to capture value without creating unmanaged risk.

Read full article

👀 Google Outlines Security Measures for Chrome’s New Agentic Features

The Story:

Google shared details on how it plans to secure Chrome’s upcoming agentic capabilities, which allow the browser to perform tasks on behalf of users, interact with websites, and automate multi-step actions.

The details:

• Chrome will restrict what agents can do by default and require explicit user approval for actions that involve data access, navigation, or changes to account settings.

• Every task an agent performs will run inside isolated sandboxes designed to prevent unauthorized access to cookies, passwords or other sensitive browser data.

• Google says agents will follow scoped permissions. They can only interact with the specific tabs, sites or data types that the user has granted.

• The browser will also include real-time monitoring to detect unexpected behavior from agents, such as interacting with domains or APIs outside the approved scope.

• Developers will have to follow new policies to ensure that third-party agents don’t request overly broad permissions or perform hidden actions.

Why it matters:

Agentic features shift the browser from a passive tool to an active participant in workflows. That creates new risks if tasks run beyond what users intended or if compromised agents gain access to private data. Clear permission boundaries and continuous monitoring will be essential as agent-based automation enters consumer and enterprise environments.

Read full article

🚨 Researchers Find 30 Security Flaws Across Major AI Platforms

The Story:

Security researchers disclosed 30 vulnerabilities affecting AI offerings from multiple vendors, including Amazon, Google, Microsoft, IBM, OpenAI and others. The issues ranged from data exposure to unsafe agent behavior.

The details:

• The flaws spanned cloud AI services, agent frameworks, plugin ecosystems and inference tools.

• Several vulnerabilities allowed unauthorized access to sensitive data or unintended API calls triggered by crafted prompts.

• Some issues stemmed from insufficient controls around agent permissions, enabling tasks to escalate beyond their intended scope.

• Other weaknesses were tied to misconfigurations in the way AI models processed inputs, which created opportunities for prompt injection or manipulation.

• Vendors have begun patching the reported problems, though researchers note that similar gaps will continue to surface as AI systems grow more complex and interconnected.

Why it matters:

The findings highlight a common pattern: many AI products inherit security assumptions from traditional software, but agentic behavior, dynamic prompts and third-party integrations introduce failure modes that older controls don’t catch. As organizations adopt AI across workflows, understanding where vulnerabilities emerge is becoming as important as traditional application security.

Read full article

🔒 UK Intelligence Warns of Rising Prompt Injection Threats

The Story:

The UK’s National Cyber Security Centre has issued a new warning about prompt injection attacks after observing a sharp rise in real-world attempts to manipulate AI systems used in government and enterprise environments.

The details:

• Officials report that attackers are increasingly embedding hidden instructions in websites, documents and user inputs to force AI models into performing unintended actions.

• These attacks can redirect models to reveal sensitive data, generate malicious outputs, or execute tasks outside approved boundaries.

• The NCSC says the risk grows as organizations integrate AI into workflows that interact with external content, making it easier for attackers to plant triggers.

• The agency released updated guidance recommending stronger input filtering, output validation, isolation of high-risk tasks and clearer limits on what actions agentic systems are allowed to perform.

• The warning stresses that prompt injection cannot be fully “patched” because the issue stems from the way models interpret language, not from a single bug.

Why it matters:

Prompt injection remains one of the most accessible and effective ways to subvert AI systems. As more businesses deploy agents that read emails, browse the web or act on user data, the gap between normal input and malicious instruction becomes harder to distinguish. Organizations should treat any external content as untrusted and put guardrails around what their AI systems can do, not just what they can generate.

Read full article

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.