BCWSC/Takeaways/AI/Summary

• For me, the Boston College Security Winter Conference was a useful snapshot of how higher-education security community thinks about risk, technology adoption, and the future of cybersecurity.
• attack speed is increasing, institutional change is slow, and risk management is almost more social than technical.
• how security community currently views AI: The sentiment felt a bit like a city police conference discussing a strange new gang in town. Imagine a loosely organized group called “the AIs” showing up, tagging buildings and occasionally causing trouble, while also claiming they want to help the community. The result is skepticism, confusion, and a desire to maintain control. That analogy roughly captures the current posture of security professionals toward AI tools. Over the course of an entire conference, the term AI agent barely came up, and discussions about prompt-injection risks were minimal. In the talk AI Lessons Learned the Hard Way (Reid Gilman), prompt injection was mentioned only briefly and largely dismissed as not a major operational concern when using AI to summarize and categorize service tickets.
• penetration testing:, Lessons from PenTesting (Kyle Enlow), reinforced similar themes. Pen testers still rely almost entirely on human expertise rather than autonomous AI tools. They use traditional human skill sets - deep understanding of networks, Active Directory, group policy, and persistence and passion in exploring systems. While AI is transforming many fields, penetration testing culture still assumes that the safest path is simply not to use AI in the environment being tested. That leaves an open question about how organizations will eventually test the security of AI-driven systems themselves.
• vendor management: The panel Managing 3rd Party Vendor Relationships (HBS's Adam Scaramella, Mike Gioia, Jeffrey Holman, Eric Boughton) highlighted how and when security teams let technologies into the organization. This role gives them influence and puts them in a difficult position as reluctant institutional gatekeepers. Their job is oriented towards slowing down adoption, insisting on or even just explaining compliance frameworks like SOC2 or HECVAT, and playing "curb your enthusiasm" therapist to excited stakeholders about why a new vendor introduces risk. In practice, conferences like this serve as a place where security professionals exchange narrative or policy strategies for maintaining control while trying to not expose the attack surface. It doesn't sound fun or easy. <!-- secretlint-disable-line -- speakers' public affiliations, not first-person -->
• data minimization: From a technical standpoint, several talks reinforced a shift in thinking about defense. In The Zero-Day Deluge: Why Patch-Centric Security Is No Longer Enough (Nathan Hall), the argument was that the window between vulnerability disclosure and exploitation is shrinking so quickly that patching alone cannot keep organizations safe. The more strategic response is data minimization - legal constraints will make it so organizations minimize and reducing the amount of sensitive information stored in systems, so that breaches cause less damage. As someone interested in AI enablement, it made me realize that we are kind of in a golden era right now - AI tools are exciting to try out and they can help us out, but it may be that the drawbridge is being pulled up for reasons beyond the scope of AI security, and soon, it will not be feasible to have AI help us out on enterprise data.
• passkeys: The talk Why Passkeys Are the Right Answer ... and Still a Nightmare (Jacob Backon and Mat Cunha) - of course passkeys are better, but it's really hard to roll-out. There are many situations where it just doesn't quite make sense yet (like ... you want to SSH to a bastion

  or jump server). For me personally, the part that's confusing about passkeys is that now our authentication is tied to something you have ... but it's often unlocked with biometrics, which is something you are. I have started using passkeys extensively both for myself and for my family, so I felt some sympathetic resonance with this struggle. The discussion included the typical confusion about "what is MFA?" and "is it two factors if a fingerprint unlocks a passkey?" - these are still non-transparent warts on the passkey narrative.
  

• Finally, the session Everything You Never Wanted to Know About Google Cloud Service Account Keys (Jordan Bradford) reinforced a simple but powerful rule: long-lived secrets should be eliminated wherever possible. Static credentials remain one of the most common ways systems are compromised. I want to get there. But the path to that promised land is riddled with IAM and OIDC and ... who knows what else. I was surprised that nobody was talking about IAM and how it needs to change for AI agents.