AI
Trusted AI as a substitute for cryptographic private inference
Ilia’s hot take: cryptography is optional if both parties trust the same model. His argument, drawn from the Yao Millionaires Problem, is straightforward. Two parties agree on a model, a prompt, and the allowed outputs. They each type in a number. The model answers who is bigger. No homomorphic encryption required.
“trusted AI models enable a new kind of private inference whereby you don’t need cryptography anymore” — Ilia
He walked through Yao’s Millionaires Problem solved with a trusted LLM (02:42) as the canonical example. The killer application, he said, is trusted execution environments where one party loads the model as a private auditor to answer questions about code they cannot hash or inspect directly.
The multi-agent gossip problem in shared workspaces
Chris Blanco, head of application security at ClickUp, named something most talks skipped: the multiplayer problem. ClickUp’s workspace puts every productivity tool in one place, so agents from different users — with different access scopes and different external integrations — all share the same environment. One agent reads a summary another agent wrote, and that information moves across permission boundaries before any check can catch it. “It’s really trivial for one bit of information to go from one agent to another faster than anyone can reasonably control it,” Blanco said. Per-agent controls catch the first hop. They miss everything after.
Using AI defensively against pig-butchering scams
Pig-butchering is a multi-billion-dollar fraud industry run today by human trafficking victims, not AI. Ariana, a data-driven security researcher and former AI skeptic, has the data to prove scammers have not yet automated it at scale. That window is closing. Once operators add AI, they can run more conversations, with better targeting, against more victims simultaneously. Her team is building AI-based defenses now, before that happens, to protect the people who will never see it coming.
Agentic red teaming as continuous mandatory defense
Rob Joyce ran NSA’s offensive hackers. His edge was never the talent or the tools. It was knowing target networks better than the defenders did. Agentic systems now do that research automatically, finding unpatched systems, shadow IT, and misconfigurations at scale.
“whether you like it or not, somebody’s going to do it to you” — Rob Joyce
His ask to every CISO in the room: run agentic red teaming on your own network continuously. The survey is coming. The only choice is who gets the findings first. One attendee is already pushing past what cloud models will do, running eight RTX 3090s on local uncensored models (13:08) to cover that last 10%.
Five-question agentic trust framework for non-technical leaders
Josh, from massivescale.ai, frames agentic AI as a digital employee. His five questions, published on the Cloud Security Alliance blog, make the risk concrete for any leader: Who are you (identity)? What are you doing (behavior)? Where are you going (segmentation)? What are you eating and serving (data governance)? What if you go rogue (anomaly detection)? Would you hand your intern the keys to the safe on day one?
“trust is earned. It’s not granted.” — Josh
Enterprise AI adoption without enterprise-grade security controls
Enterprises are deploying GenAI at breakneck speed. The security controls are not keeping pace. Zraini Ramolingum from Snowflake put it plainly:
“security controls are an afterthought” — Zraini Ramolingum
Features ship first. Security gets negotiated in afterward, vendor by vendor, deal by deal. The ask: the security community presses AI vendors to bake enterprise-grade controls in at release, not retrofit them under pressure.
Notable Quotes
trusted AI models enable a new kind of private inference whereby you don’t need cryptography anymore Ilia · ▶ 2:08
we are absolutely not ready for this for what’s happening right now Greg Notch · ▶ 4:38
whether you like it or not, somebody’s going to do it to you Rob Joyce · ▶ 11:53
trust is earned. It’s not granted. Josh · ▶ 15:36
security controls are an afterthought Zraini Ramolingum · ▶ 17:14
Key Takeaways
- Run agentic red teaming on your own network continuously — attackers already are.
- Treat agents as digital employees and apply five identity/behavior/governance questions before granting access.
- Pig-butchering fraud has not yet adopted AI at scale; defenders must build countermeasures now, before it does.