Future of Data Security

George Al-Koura refuses to let AI agents run in his production environment. As CISO at ruby, the parent company of Ashley Madison, he's protecting data where a breach doesn't just expose PII but reveals people's most private thoughts and relationships across a global user base. George tells Jean why the hardest data security challenge is still foundational: too many leaders in the space can't distinguish structured from unstructured data, and organizations keep throwing agents at the problem without understanding the manual processes they're trying to automate, which is exactly why they're not seeing ROI on their AI spend.

George is also pitching the Canadian federal government on a concept he calls the AI Training Data Bill of Material (TDI BOM), modeled after SBOMs: a compliance process that produces a verifiable report ensuring the provenance of data used to train models. He cites studies showing that corrupting less than half of 1% of a model's training data can compromise the entire model, and if that model runs targeting data for defense systems or critical infrastructure like water treatment, the failure mode goes well past data loss. He's pushing for TDI BOMs to be required in government procurement, starting with critical infrastructure supply chains, as a step toward digital and data sovereignty. On the commercial side, George co-founded Very Data Free, a veteran-founded secure-by-design platform he describes as "eBay for your data," built to let organizations sell or loan proprietary datasets for AI model training. The conversation also covers how the SIEM-era centralized security model was built for log aggregation and breaks down at petabyte-scale file data, and why GenAI is forcing organizations to finally secure unstructured data environments they've been ignoring.

Topics discussed:
Refusing to let AI agents access production logs and environments

AI Training Data Bill of Material as a government procurement requirement

Model poisoning risks at sub-0.5% training data corruption thresholds

Mapping manual processes before AI automation to prove ROI

Centralized SIEM-era architecture failing at petabyte-scale unstructured data

GenAI forcing organizations to secure previously ignored file environments

AI-generated fake passports and government IDs bypassing identity verification

Hiring self-taught operators over certification-heavy candidates for SOC teams

Kate Helin, Legal Director of Privacy & Data Security at Snyk, argues that agents have already become the biggest security risk in most enterprise tech stacks, and that most organizations are not set up to address it. The core problem is not a lack of controls. It is that no single function has full visibility into how agents behave. Kate's approach is to convene legal, security, R&D, and GRC before any mitigation decision is made, because legal cannot counsel on obligations until the technical teams explain how the technology actually works. The composition of that conversation determines whether the resulting control is technical, human, or both.
Kate also draws a direct line from GDPR implementation to today's AI governance challenges. She describes how building privacy programs under early GDPR, when implementation details were absent and community norms had to substitute for regulatory guidance, prepared her to operate in the same conditions now present in AI. Her operating principle is to meet the spirit of the law when the prescriptive details have not been written yet.
Topics discussed:
Why agentic AI has become the biggest current security risk across most enterprise tech stacks

Structuring cross-functional roundtables across legal, security, R&D, and GRC before agentic risk controls are selected

How early GDPR implementation under regulatory ambiguity prepared privacy counsel for today's AI governance challenges

Applying the spirit of the law when prescriptive AI regulation has not yet been written or enforced

Why technology consistently outpaces regulation and what that means for security teams building compliant programs today

Using AI as a distillation tool for complex legal and security analysis while maintaining human-in-the-loop validation

Why junior lawyers and engineers still need mentorship to develop judgment that AI-generated outputs cannot replace

Stephen Fridakis, CISO in Residence at Cyderes, comes to this conversation with a framework that cuts against how most security teams still operate: stop thinking about perimeters, start thinking about consequences. His argument is that the question of "are we secure or not" is not just unhelpful, it's the wrong unit of measurement entirely, and he offers a more honest alternative built around what an organization can afford to lose versus what must never leave.
Stephen makes a precise and underappreciated case for why shadow AI is fundamentally different from every other control problem a CISO has faced. Once sensitive data is submitted to a public model, it is embedded, transformed, and learned. There is no rollback. The most effective response is not detection after the fact but building organizational awareness before the decision to submit is ever made. He also breaks down why static trust models have collapsed under AI, arguing that just-in-time data access and ephemeral credentials are no longer aspirational, they are necessary, and why past behavior can no longer serve as a proxy for future safety.
Topics discussed:
Reframing CISO governance around consequence management rather than perimeter defense or binary secure/not-secure assessments

Applying the afford-to-lose framework to prioritize finite security budgets against the data that matters most

Understanding AI irreversibility as a distinct control problem where sensitive data submitted to public models cannot be retrieved

Shifting shadow AI strategy from post-submission detection to pre-decision awareness building across the organization

Replacing static role-based trust models with context-driven identity evaluation that accounts for data stage and purpose

Moving toward ephemeral credentials and just-in-time data access as the foundation of modern security architecture

Evaluating where AI delivers real operational value versus where uncontrolled use produces unreliable and unexplainable outputs

Advising new CISOs to build both technical depth and business fluency to avoid the most common leadership failure points

TELUS didn't wait for generative AI to arrive before building governance infrastructure. Jesslyn Dymond, Director of AI Governance & Data Ethics, joined the company in 2019 to stand up responsible AI practices alongside the machine learning teams building them, which meant that when generative AI hit, the governance scaffolding was already there. Jesslyn walks through the specific structures TELUS uses to govern AI at scale: a CEO-led AI board that includes the CIO, Chief AI Officer, and Chief Data and Trust Officer; a network of hundreds of data stewards embedded across business units and appointed by VPs; and a unified intake process called a Data Enablement Plan that consolidates privacy, security, and responsible AI review into a single workflow instead of separate forms and sign-offs.
Jesslyn also shares how TELUS certified its first generative AI customer support tool to the international Privacy by Design standard and then had it independently audited, and what that process required the team to work through on transparency and user experience. She makes a pointed case for why shadow AI is best addressed with access to better internal tools rather than policy restriction alone, explains how her team grades levels of agency within their agentic AI framework to determine what controls need to be in place before approving systems, and describes how TELUS took the concept of purple teaming out of the security world and applied it to AI governance, including running those sessions with students and the general public.
Topics discussed:
Building proactive AI governance infrastructure before adoption by embedding responsible AI practices alongside ML development teams

Structuring enterprise AI oversight through a CEO-led board including CIO, Chief AI Officer, and Chief Data and Trust Officer

Deploying VP-appointed data stewards across business units to connect governance policy with on-the-ground AI implementation

Consolidating privacy, security, and responsible AI review into a single Data Enablement Plan to reduce friction and improve compliance 

Certifying a generative AI customer support tool to the international Privacy by Design standard and navigating external audit requirements

Grading levels of agency within an agentic AI framework to determine appropriate controls

Countering shadow AI by prioritizing internal tool access and functionality over policy restriction alone

Applying purple teaming from security practice to AI governance to test systems collaboratively across various teams

Polymer's runtime security approach operates at the file and message level, intercepting content in real-time within workflows like Slack and Zendesk to redact, block, or grant granular access based on specific entities found inside documents. This contrasts with traditional perimeter-based security where access is binary: you're either in the club or out. Yasir Ali, Founder & CEO of PolymerHQ DLP, explains how financial services has operated under workflow-level distrust for over a decade, with every file interaction requiring labeling and ethical wall policies between trading and investment banking divisions, and why the rest of the enterprise world is finally moving toward this model.
Yasir also touches on a critical gap in current security architectures: control planes across network, identity, and content layers don't communicate with each other. His team works to triangulate telemetric data from tools like Zscaler with Polymer's ground-level content controls, creating unified policy layers without forcing organizations into single-vendor platforms. He also addresses a tension in AI-powered security: probabilistic detection models work well for entity recognition, but policy enforcement must remain deterministic. You can't have AI deciding some days to block sensitive data and other days letting it through.
Topics discussed:
Implementing runtime security at file and message level to enable partial document sharing based on entity-level access policies

Solving the binary sharing problem in unstructured datasets where traditional security forces all-or-nothing file access 

Adopting financial services workflow-level distrust model that requires labeling and ethical wall policies for all file interactions

Addressing enterprise AI adoption barriers through proper identity modeling for non-human agents and machine-to-machine interactions within IAM systems

Triangulating telemetric data across network, identity, and content control planes to create unified policy layers without vendor lock-in

Balancing probabilistic AI detection models for entity recognition with deterministic policy enforcement to maintain response certainty

Building enterprise software teams by prioritizing cultural fit and collaboration ability over hiring 10x engineers