RSCA – Day 4 Reporting
RSAC 2026 – Day 4 Report (March 26, 2026) AUDIO
AI Is Moving Into Production Faster Than Security Can Keep Up
By Jorge Avila, KCXU 92.7 FM Reporter – Cyber / AI News Desk
SAN FRANCISCO — Day 4 of the RSAC 2026 Conference made one thing unmistakably clear:
Artificial intelligence is no longer sitting in the lab. It is now inside production systems, developer workflows, business operations, and critical infrastructure.
And according to many of today’s sessions, that shift is happening faster than most organizations are prepared to secure.
From AI agents acting with too much autonomy, to AI-assisted coding introducing hidden vulnerabilities, to deepfakes, healthcare incident response, and the hardening of AI infrastructure for national security, Day 4 was less about AI hype and more about AI operational reality.
Day 4 Theme: AI Is Now an Operational Security Problem
If Day 1 was about strategy, Day 2 about technical disruption, and Day 3 about community impact, Day 4 was about implementation risk.
Across multiple sessions, one message kept surfacing:
Organizations are already deploying AI into real workflows before governance, architecture, and security controls have fully caught up.
One of the clearest examples came from the session “Is Your AI Agent a $10M Asset or a $10M Liability?”, which pointed to a widening gap between adoption and oversight:
● 79% of organizations are already using AI agents
● But only a small fraction have meaningful security governance around them
That mismatch was one of the strongest signals from Day 4.
The AI era is no longer about asking “Should we use it?”
It is now about asking:
“How do we keep it from becoming our next major attack surface?”
AI Agents: Powerful, Productive, and Potentially Dangerous
One of the biggest Day 4 storylines was the rapid rise of AI agent systems that do more than generate text. These tools can now:
● retrieve information
● write files
● call APIs
● automate tasks
● interact with business systems
● and in some cases, act semi-autonomously
That sounds powerful. But it also introduces a serious new risk model.
The IANS material used in one RSAC session laid out five high-level agentic AI risks:
● excessive autonomy
● privilege escalation
● sensitive data disclosure
● goal manipulation attacks
● lack of accountability
And the explanation was simple but effective:
AI agents can behave less like reliable software and more like a very literal assistant with too much access.
That matters because once an AI agent is given:
● access to email
● cloud systems
● documents
● internal tools
● or automation rights
…it can create real business risk even without malicious intent.
“Vibe Coding” and AI-Assisted Development Are Raising Security Concerns
Another major focus on Day 4 was AI-assisted coding, sometimes called “vibe coding,” where developers use AI tools to rapidly generate or modify software.
The session “Insecure Vibes: Avoiding the Risks of AI-Assisted Coding” made a blunt point:
AI coding tools often learn from mostly public code, with mixed quality, where security is optional in the data
That means these systems can produce code that is:
● syntactically correct
● fast to generate
● but not necessarily secure
And if developers over-trust those outputs, the result can be faster software delivery paired with higher security debt.
The recommended response was not to reject AI coding entirely but to wrap it with guardrails, including:
● secure coding standards
● privacy requirements
● security review tooling
● secure SDLC practices
● and training developers to critically evaluate AI-generated code
That is a key Day 4 takeaway:
AI can speed up development, but it can also speed up insecure development if used carelessly.
AI Security Is Bigger Than Red Teaming
Another strong theme from today: Red teaming alone is not enough anymore.
A session titled “Beyond Red Teaming: Why AI Security Needs a Bigger Playbook” argued that organizations need a much broader security model for AI, including:
● AI red teaming
● RAG defenses
● supply chain governance
● runtime monitoring
● response and recovery
That reflects a broader shift happening across RSAC this year:
Security leaders are realizing that AI cannot be protected with only traditional application security methods.
It now requires:
● governance
● monitoring
● testing
● architecture controls
● and incident response playbooks built specifically for AI systems
Real-World Agent Exploits Are Already Here
One of the more alarming technical examples from Day 4 came from the session “When Your AI Agent Works for Me.”
That presentation showed how AI coding and agent systems can be manipulated into bypassing user protections and modifying sensitive files.
One example showed how a path case-sensitivity issue could be abused to write into sensitive locations and modify configuration files without expected user approval. The session also documented scenarios involving auto-approved tool calling and command execution risks in agent workflows
That matters because it moves the AI security conversation from theory to proof:
These systems are already creating new exploit paths.
And in many organizations, they are being introduced into environments that were never originally designed for autonomous or semi-autonomous software actors.
Healthcare Cybersecurity: A Reminder That Security Is Also a Human Safety Issue
Another important Day 4 session focused on proactive cyber incident management in healthcare, and it served as a reminder that cybersecurity failures do not just impact data.
They can affect:
● care delivery
● trust
● operations
● compliance
● and public safety
The healthcare-focused material highlighted recurring incident themes such as:
● human-factor vulnerabilities
● policy gaps
● confidentiality and regulatory breaches
● reputational damage
● and weak visibility over sensitive data handling
The proposed lessons were practical and highly relevant:
● improve traceability
● strengthen awareness training
● formalize external data-sharing rules
● and improve detection through log correlation and audit tooling
That’s a valuable Day 4 reminder:
Cybersecurity is not just an IT issue in sectors like healthcare; it is directly tied to continuity, trust, and human well-being.
Deepfakes and Adversarial AI Continue to Expand the Threat Landscape
Another Day 4 track looked at the rising role of deepfakes and adversarial AI in cybersecurity.
The session “Defending Against Adversarial AI & Deepfake Attacks” framed the issue clearly:
● AI threats are no longer limited to traditional exploits
● Deepfakes are distorting trust itself
● And social engineering remains the preferred delivery mechanism for AI-enabled attacks
This is especially important because deepfakes are no longer fringe tools. They are increasingly plausible, scalable, and usable for:
● impersonation
● extortion
● fraud
● support scams
● internal business deception
● and disinformation operations
That reinforces one of the broader RSAC 2026 themes:
In the AI era, trust itself is becoming a security boundary.
AI Infrastructure and National Security Are Becoming Tightly Linked
Another notable Day 4 topic was the session “From GPU to Grid: Hardening AI Infrastructure for National Security.”
Even from the session framing alone, the message was significant:
AI is no longer just a software or enterprise issue. It is now a matter of:
● compute infrastructure
● resilience
● power dependency
● supply chain integrity
● and national-level strategic security
That matters because AI capability increasingly depends on real-world infrastructure:
● GPUs
● data centers
● networking
● energy reliability
● and hardened supporting systems
And if that infrastructure becomes vulnerable, the downstream consequences go far beyond individual companies.
Governance Is Becoming the Difference Between Innovation and Exposure
A quieter but highly practical Day 4 theme was AI governance, especially for smaller organizations.
The session “AI Governance for SMEs: Innovate. Sell. Secure.” underscored a growing reality:
Small and mid-sized organizations do not get a pass on AI risk.
In fact, many are under pressure to:
● adopt AI quickly
● Demonstrate trust in customers
● and do so with fewer legal, privacy, and security resources than larger enterprises
That matters because many organizations in local communities, including small businesses, service providers, clinics, and nonprofits, fit that profile exactly.
And for them, AI governance is not about bureaucracy.
It is about:
● trust
● safety
● privacy
● and business survivability
Why This Matters to KCXU / Our Community
This was one of the strongest “why this matters” days of the conference.
Because Day 4 showed that AI risk is no longer confined to:
● advanced research labs
● Big tech companies
● or elite cybersecurity teams
It is now moving directly into:
● The software people use at work
● The systems that store community data
● The tools that small businesses may adopt without much scrutiny
● The healthcare institutions families depend on
● and the communication channels people trust
That matters deeply for KCXU’s audience.
Because many underserved communities, small organizations, and local institutions do not have:
● dedicated AI governance teams
● internal red teams
● mature AppSec programs
● or staff who can evaluate whether an AI tool is safe before it is adopted
And that creates a real gap.
A gap between:
● AI capability
● and AI readiness
And if that gap is not addressed early, communities do not just miss the upside of AI.
They absorb the downside.
That means:
● more exposure to fraud
● more insecure systems
● more over-trusted automation
● and more risk is pushed onto organizations that are least equipped to absorb failure
That is exactly why KCXU’s IT / AI / Cyber coverage matters.
Because the public conversation around AI often focuses on:
● speed
● convenience
● disruption
● and opportunity
But what Day 4 at RSAC showed is this:
Security, governance, and resilience must move with that same urgency — or the cost will be paid later.
Final Day 4 Takeaway
Day 4 of RSAC 2026 delivered one of the clearest messages of the week:
AI is no longer a future security challenge. It is a present operational one.
It is being deployed now. Integrated now. Trusted now. And in many places, it is not yet governed well enough.
The organizations and communities that will do best in this next phase will not be the ones that adopt AI the fastest.
They will be the ones who adopt it:
● intentionally
● securely
● with guardrails
● and with a clear understanding of the risks
By Jorge Avila KCXU 92.7 FM Reporter Cyber / AI News Desk