RSAC 2025: AI is Everywhere. Trust? Not So Much.

Just wrapped up a packed, somewhat frenetic, but mostly enjoyable RSAC 2025 Conference. And if I had to sum it up in a sentence: AI is everywhere, but trust and control are still catching up.

The conference opened with a shot fired across the bow of the security and identity industry. Patrick Opet, CISO of JPMorganChase, published an open letter challenging the lack of secure-by-design thinking in modern integration patterns within the SaaS world, which is breaking down essential security guardrails. He specifically called out how the success of modern identity protocols (like OAuth) in enabling these integrations is actually leading to complacency in the pursuit of speed and convenience, and that improper setup of permission scopes and controls is creating vulnerabilities across a far broader and distributed attack surface.

In a way, his call-to-action both complements and contradicts what I saw at RSAC 2025. AI is speeding its way into the IT and SOC infrastructure for every organization – not as a theoretical capability, but as a triage partner, copilot, and even autonomous investigator (with decision maker just around the corner). When it comes to the SOC, however, it’s becoming clear that flashy dashboards aren’t enough any more. CISOs and practitioners are looking for proof that these tools drive outcomes at scale and don’t just shift complexity elsewhere. You can’t just sprinkle some AI on your existing offerings and call it innovation.

Beyond tooling, a deeper theme emerged: AI is the new operational surface and the new attack surface. From agent memory manipulation to prompt injection, organizations are discovering vulnerabilities that didn’t exist a year ago. And with AI wrappers, SaaS sprawl, and loosely governed agent identities, enterprise risk is evolving faster than our control models.

Here’s what stuck with me most:

• Identity is fragmenting fast: Humans, bots, APIs, and AI agents now live in parallel – each with its own lifecycle, permissions, and risks. Traditional IAM isn’t cutting it. Identity Security was one of the main themes, but few people outside of the identity bubble can properly define it. NHIs are taking over the world (hopefully not in the literal sense). Folks by and large understand that identity is key, and are paying increased attention, especially to delegated authorization, agent-specific policy enforcement, and fine-grained data access controls.
• We’re not there yet on passkeys: While the news on passkey adoption and rollout continues to be encouraging, discussions at the conference show that it isn’t quite breaking through yet. Usability is still a major concern, manageability is viewed as a challenge, and people not immersed in identity still have fundamental questions. World Passwordkey Day happened to fall during the conference, but it is clear that there is still a lot of (hard) work to do to overcome the hesitation many stakeholders still have.
• Memory is now a risk vector: Persistent agents that “learn” from interactions can also be misled by them. Attackers will go after context the way they once went after credentials. We need to secure not just the data agents access, but the logic and memory they build from it.
• Identity and security must enable, not obstruct: Too many orgs still see security as something to bolt on – or worse, slow things down. But the smart ones are building it into the architecture from day one, recognizing that sacrificing a little bit of speed for control and resilience is more efficient in the long run. Identity Security has a critical role to play in this area, in making security and enablement scale together.
• The cost curve needs a reality check: Using LLMs and AI isn’t cheap. A thousand alerts a day could cost millions in processing alone. AI-powered doesn’t always mean ROI-powered, and I was surprised that few of the customers (prospective or existing) of security tools were asking how their costs or subscription would increase due to the addition of AI powered features. We need to get sharper about value, not just capability.
• CISOs are tired. And worried: Patrick Opet isn’t alone. Everyone’s chasing context and control. While many tools are claiming to combat alert fatigue, that is viewed largely as a temporary fix. Beyond visibility, they want clarity, resilience, and the ability to scale good decisions fast. Their overarching need is to shift the focus from detect-and-respond to prevent-and-prove.

RSAC confirmed what many of us already felt: despite how we may feel about its current state, we have to accept that AI is becoming a business mandate. But without smart controls, agent-level governance, and identity models that match this new reality, we’re flying blind. We’re not just securing users anymore – we’re securing decision-making. It’s time to stop asking whether AI can do it, and start asking whether we’re building the trust infrastructure to let it.