Phil Venables has a message for the cybersecurity industry, and it isn’t gentle. Google Cloud’s Chief Information Security Officer, writing in the wake of RSA Conference 2025, argues that the profession’s approach to artificial intelligence security, talent development, and threat management needs to be rebuilt from the ground up—not incrementally improved. His assessment, published on the Google Cloud Blog, amounts to a manifesto for an industry standing at a crossroads it didn’t fully anticipate arriving at this quickly.
The timing matters. RSAC 2025, held in late April in San Francisco, drew more than 40,000 attendees and featured over 600 sessions. AI dominated the conversation—not as a novelty, but as an operational reality that’s already reshaping how organizations defend themselves and how adversaries attack them. Venables used his post-conference column to synthesize what he sees as the three defining challenges: securing AI systems themselves, using AI to transform security operations, and rethinking who does the work of cybersecurity and how they’re prepared for it.
Start with the AI security problem. Venables makes a distinction that too many organizations still blur: there’s a difference between using AI for security and securing AI. The former is well underway. Security operations centers are deploying large language models to accelerate threat detection, automate incident triage, and generate response playbooks. Google’s own Gemini models are embedded across its security product line, and competitors from Microsoft to CrowdStrike to Palo Alto Networks are racing to do the same.
But securing AI? That’s harder. And less glamorous.
Venables points to the growing attack surface that AI systems introduce—model poisoning, prompt injection, data exfiltration through inference, adversarial inputs designed to manipulate outputs. These aren’t theoretical risks. They’re being actively exploited. A Google Cloud Blog analysis accompanying the RSAC recap notes that organizations deploying generative AI models often fail to apply the same rigor to AI pipelines that they do to traditional software supply chains. Training data provenance, model integrity verification, and runtime monitoring remain afterthoughts at many enterprises.
This gap is widening, not closing. As companies rush to deploy AI agents—autonomous systems that can take actions, not just generate text—the stakes escalate. An AI agent with access to production databases, customer records, or financial systems doesn’t just answer questions. It acts. And if it’s compromised, it acts on behalf of an attacker with whatever permissions it’s been granted.
Venables argues that the industry needs to treat AI systems as first-class citizens in security architecture. Not bolted-on afterthoughts. Not experimental sandboxes disconnected from governance frameworks. Full integration into identity management, access control, monitoring, and incident response. Google has been pushing its Secure AI Framework (SAIF) since mid-2023 as a blueprint for this, and Venables used the RSAC moment to reiterate that the framework needs broader adoption beyond Google’s own walls.
The workforce question is equally urgent and arguably more vexing. Cybersecurity has been dealing with a talent shortage for over a decade. The numbers are familiar to anyone in the field: estimates from ISC2 put the global gap at roughly 4 million unfilled positions. What’s changing, Venables contends, is not just the size of the gap but its nature.
AI is redefining what security professionals need to know. The traditional career path—help desk to SOC analyst to incident responder to architect—still exists, but it’s no longer sufficient. Security teams now need people who understand machine learning operations, data engineering, model evaluation, and the specific threat models that apply to AI systems. These skills don’t come from a CISSP exam. They don’t come from most university programs, either.
Venables calls for what he describes as a fundamental rethinking of security education and career development. He’s not alone. At RSAC 2025, multiple sessions addressed the collision between AI advancement and workforce readiness, with speakers from CISA, NIST, and major financial institutions all flagging the same concern: the tools are outpacing the people trained to operate them responsibly.
There’s a paradox here. AI could help close the talent gap by automating lower-level tasks and enabling junior analysts to perform at higher levels. Google has demonstrated this with its own security operations tools, where Gemini-powered assistants can translate complex queries into natural language explanations, suggest investigation steps, and draft remediation plans. In theory, this means fewer experienced analysts are needed for routine work, freeing them for higher-order problems.
In practice, it’s messier. Automation doesn’t eliminate the need for expertise—it shifts it. Someone still has to validate the AI’s recommendations. Someone has to recognize when the model is hallucinating, when it’s been fed adversarial data, when its confidence score masks genuine uncertainty. That someone needs deep technical knowledge and the judgment that comes from experience. So the workforce problem doesn’t disappear. It transforms.
And there’s a risk that organizations will use AI as an excuse to underinvest in people. Why hire three more analysts when the AI can handle the volume? This logic is tempting and dangerous. Venables doesn’t say this explicitly, but the implication runs through his analysis: AI augments human capability, and treating it as a replacement invites catastrophic blind spots.
The geopolitical dimension adds another layer of complexity. Google’s Threat Intelligence Group, which Venables references, has been tracking state-sponsored actors—particularly from China, Russia, North Korea, and Iran—who are actively incorporating AI into their offensive operations. These groups are using generative AI for reconnaissance, social engineering, malware development, and vulnerability research. The asymmetry is stark: nation-states can invest heavily in AI-powered offensive capabilities while most defenders are still figuring out procurement.
Recent reporting reinforces this urgency. Multiple outlets covering RSAC 2025 noted that AI-generated phishing campaigns have become significantly more sophisticated in the past twelve months, with language quality and personalization reaching levels that make traditional email security filters less effective. The volume is increasing too. Attackers can now generate thousands of unique, contextually tailored phishing messages in minutes.
Venables also touches on a theme that resonated throughout the conference: the convergence of cloud security and AI security. As most AI workloads run in cloud environments, the security of the underlying infrastructure becomes inseparable from the security of the models themselves. Misconfigurations in cloud storage can expose training data. Inadequate identity controls can let unauthorized users access model endpoints. Weak logging means compromises go undetected.
Google, naturally, has a commercial interest in this argument. The company’s cloud security portfolio—spanning Chronicle, Mandiant, and the broader Google Cloud security suite—is positioned to address exactly these converged challenges. But the underlying point holds regardless of vendor: organizations can’t think about AI security in isolation from their cloud security posture. The two are entangled.
One of the more provocative elements of Venables’s analysis concerns the role of regulation. He acknowledges that governments worldwide are moving to regulate AI, from the EU AI Act to various U.S. executive orders and state-level proposals. His position, consistent with Google’s broader stance, is that regulation should be risk-based and focused on outcomes rather than prescriptive technical mandates. He warns that overly rigid rules could stifle innovation and create compliance burdens that don’t actually improve security.
This is a familiar argument from Big Tech, and it deserves scrutiny. The counterpoint—made by numerous RSAC speakers and in publications like Wired and MIT Technology Review—is that voluntary frameworks and industry self-governance have a mixed track record at best. The tension between innovation speed and regulatory rigor isn’t going away. If anything, the pace of AI development is making it more acute.
What Venables gets right, and what makes his post-RSAC analysis worth serious attention, is the framing. He doesn’t treat AI security as a niche subdiscipline or an emerging concern for future consideration. He treats it as the central organizing challenge of modern cybersecurity. Everything else—cloud security, identity management, threat intelligence, workforce development, regulatory compliance—connects to it.
That’s a significant shift from even two years ago, when AI security was one topic among many at conferences like RSAC. Now it’s the topic. And the organizations that fail to recognize this, Venables suggests, will find themselves defending against 2025’s threats with 2019’s assumptions.
The practical implications are substantial. CISOs reading Venables’s analysis should be asking hard questions. Does our security architecture account for AI-specific threats? Do we have visibility into our AI supply chain—the models, the training data, the inference pipelines? Are our security teams equipped to evaluate AI systems, or are we relying on data science teams who may not think in security terms? Are we building career paths that prepare the next generation of defenders for a world where AI is both the tool and the target?
These aren’t abstract questions. They’re operational ones. And the window for treating them as someone else’s problem is closing fast.
Google’s own investments reflect the urgency. The company announced several security-related AI capabilities at Google Cloud Next in April 2025, including enhanced threat detection powered by Gemini, automated security posture management, and expanded Mandiant services for AI incident response. These moves signal that Google sees AI security not just as a product category but as a competitive differentiator in the cloud wars.
Microsoft is making parallel bets with its Security Copilot platform. Amazon Web Services is integrating AI into GuardDuty and other security services. The hyperscalers are all converging on the same thesis: the future of cloud security is AI-native, and the vendors who get there first will capture disproportionate market share.
But vendor solutions alone won’t solve the problem. Venables is clear on this point. Technology is necessary but insufficient. The human element—training, judgment, organizational culture, cross-functional collaboration between security teams and AI teams—remains decisive. The best AI-powered security tools in the world are useless if the people operating them don’t understand what they’re looking at.
So where does this leave the industry? In a position that’s simultaneously more capable and more vulnerable than it’s ever been. The tools are better. The threats are worse. The talent pipeline is misaligned. And the pace of change isn’t slowing down for anyone.
Venables ends his analysis on a characteristically direct note: the time for pilot programs and proof-of-concept exercises is over. AI security needs to be embedded in production environments, governance frameworks, and hiring strategies now. Not next quarter. Not after the next board meeting.
Now.
Whether the industry listens is another matter entirely. But the argument has been made, clearly and publicly, by one of the most senior security executives at one of the world’s most influential technology companies. The excuse of not knowing is no longer available.
WebProNews is an iEntry Publication