As enterprises rush agentic AI into production, the security conversation is shifting from perimeter defence to behavioural trust. In this conversation with Emilio Escobar, Chief Information Security Officer at Datadog, he unpacks why agent compromise rarely trips an alarm, why dashboards stay green even as an agent operates beyond its intended scope and why least-privilege identity for agents can no longer be an afterthought.
The discussion also turns to India-specific pressures: how architectures built around data sovereignty are reframing DPDP Act compliance from a policy promise into a verifiable fact, and why regulatory portability across CERT-In, RBI, and global frameworks starts with the strictest common-denominator control set, not parallel compliance stacks.
From vendor trust reciprocity to board-level risk translation and the non-delegable core of the CISO mandate, this serves as a wide-ranging look at what security leadership demands as automation reshapes both the threat landscape and the talent pipeline beneath it.
CISO Forum: AI Guard / agentic attack surface: What’s the minimum security baseline before granting agents any production access?
Emilio Escobar: A security baseline should comprise three controls. First, a centralised, continuously updated inventory of every agent in operation, its authorised scope, and the data it touches. Without this, defence is structurally impossible. Second, least-privilege identity architecture applied to agents with the same rigour as service accounts, avoiding standing broad access granted for convenience. Third, behavioural monitoring at the action level rather than the request level, since a compromised agent typically doesn’t error out; it continues executing, invoking tools and querying systems in ways indistinguishable from normal operation until the behavioural pattern is examined.
This practice helps catch an incident in minutes, not days later. The core takeaway is that an AI agent can operate beyond its intended purpose while every dashboard stays green, because traditional metrics track uptime and request completion — not whether the agent is doing what it is supposed to do.
CISO Forum: Data sovereignty and Bring Your Own Cloud: How does this change the compliance conversation under India’s DPDP Act?
Emilio Escobar: Under the DPDP Act, residency and processing-location scrutiny require organisations to demonstrate, not merely assert, where regulated data is processed and stored. When customer data remains in the customer’s own storage layer throughout, the cross-border processing question is structurally eliminated rather than mitigated through policy commitments, materially simplifying compliance review for regulated sectors such as BFSI and GCC operations. Datadog’s Bring Your Own Cloud (BYOC) architecture shifts data residency from a contractual assurance to a verifiable technical fact.
CISO Forum: Vendor trust reciprocity: What obligations does a platform vendor owe, and what should Indian CISOs be asking that they currently aren’t?
Emilio Escobar: A vendor with infrastructure-level access and log custody owes customers the same standard of transparency it expects of its own environments: current, verifiable compliance posture and a clear technical delineation of what the vendor’s platform can observe and act on inside any environment. Datadog’s privacy program is specifically designed around these principles: transparency, privacy by design, and global compliance. This means customers should control what data reaches the platform, rather than relying solely on vendors to manage their sensitive data. Subprocessor changes should be disclosed with enough lead time to exercise the right to object, not just enough to satisfy a contract clause that nobody reads until an incident occurs. And lastly, vendors need to align with the customer’s jurisdiction, including breach timelines, consent requirements, grievance mechanisms, and data principal rights.
Most Indian CISOs stop at the certification checklist. The core question is whether the vendor’s breach notification commitment is written to its own internal SLA or to CERT-In’s six-hour clock, whether post-termination data deletion is something they can verify or only something they’re told, and whether their telemetry or logs feed any AI model training with no per-workspace opt-out. Those questions test whether a vendor’s practice matches its posture.
CISO Forum: Detection-to-remediation economics: More automation, or more human judgment, for resource-constrained teams?
Emilio Escobar: For teams with limited security resources, it’s not a question of one or the other. Instead, it’s a matter of automating functions – such as prioritisation and correlation – so their people can focus on strategic initiatives that matter most. Datadog’s State of DevSecOps 2026 report found that applying runtime context allows up to 97% of vulnerabilities flagged “critical” by static severity scoring to be deprioritised because they are not actually reachable in production. For a resource-constrained team, the highest-leverage investment can be automated triage that directs scarce analyst time to genuinely exploitable risk rather than static-severity noise. This does not remove human judgment from the process; it makes that judgment economically viable at scale.
CISO Forum: Observability-security organisational convergence: Should CISOs push for org convergence, or is tool convergence sufficient?
Emilio Escobar: Tool convergence alone is insufficient if security and engineering functions remain organizationally separate, with distinct reporting lines, incentives, and on-call structures. The handoff delay commonly cited as a detection failure is more often an accountability gap, a finding crossing an organisational boundary before ownership for remediation is assigned.
Security and risk management are a team sport. A full organisational merger isn’t necessary for every enterprise, but shared incident ownership and shared operational metrics between security and SRE functions should be the target state, regardless of reporting-line structure.
The cost of getting this wrong isn’t abstract; it shows up in resolution speed. The core point is simple: a unified toolset that surfaces a threat in seconds delivers little value if that finding then sits for hours or days waiting for someone to claim it, because the security team that spotted it and the engineering team that can fix it were never measured against the same outcome. That gap is exactly where a detected issue becomes an incident, and where a contained incident becomes a disclosed one. CISOs don’t need to redraw the org chart to close it, but they do need to insist that “who owns this the moment it’s found” is answered before the tooling is, not after.
CISO Forum: Regulatory portability: What patterns satisfy multiple regulatory regimes, and does this translate to CERT-In, RBI, and India’s sector mandates?
Emilio Escobar: When a security architecture is built for one regulatory regime, the question of whether it will also satisfy the next one is usually the wrong question—the more useful pattern is architecting to the strictest common-denominator control set first: continuous monitoring, comprehensive audit logging, and verifiable access governance. From there, an organisation can map a single architecture to each framework’s specific requirements, rather than maintaining separate, parallel compliance stacks for every regime it operates under.
Datadog’s progression from FedRAMP Moderate to FedRAMP High followed this model. The same underlying disciplines—continuous verification of data control and provenance—map directly onto CERT-In’s incident reporting timelines and RBI’s data localisation and audit expectations, both of which ultimately test the same architectural property through different regulatory language.
CISO Forum: Talent and the “usable security” philosophy: How should hiring and upskilling change over the next 3–5 years?
Emilio Escobar: As automation absorbs more of the tool-specific, alert-queue work that has traditionally defined entry-level SOC roles, hiring and development should shift earlier toward investigative reasoning and systems thinking—the ability to synthesise a login anomaly, a traffic spike, and a configuration change into a coherent incident narrative, rather than familiarity with any single alert queue or tool.
Existing analyst capacity should be redirected accordingly—toward the judgment-intensive work that automation can support through accurate escalation but cannot fully resolve, such as acceptable-risk determinations in ambiguous cases, where the right call depends on business context automation doesn’t have. For security professionals building careers across India’s technology and BFSI sectors, the core takeaway is to prioritise durable, non-automatable skills now: not because tool fluency stops mattering, but because tool fluency has a shrinking half-life, while the ability to reason across a system under pressure does not.
CISO Forum: Board-level risk translation: How should CISOs report risk when boards may not distinguish system-detected from human-detected incidents?
Emilio Escobar: Board reporting should focus on business outcomes — incidents prevented and the cost of downtime — rather than which tool or layer detected a threat. Whether a human or an autonomous agent takes action is an internal accountability question; ownership of the outcome ultimately remains with the CISO.
Boards also need to understand a broader misconception: a threat is no longer always a discrete event with a clear alarm. Modern attacks often emerge as subtle patterns across signals — a login anomaly, performance dip, or unexplained uptime shift. Context matters too. As stated earlier, most vulnerabilities that appear urgent on paper may not reflect the real risk.
That is why outcome-based reporting is more than a communication tool. It gives boards a clearer view of operational risk by focusing on what affects the business, rather than raw severity labels or detection volumes.
CISO Forum: The future CISO mandate: What is the CISO’s non-delegable responsibility?
Emilio Escobar: What remains non-delegable is the determination of organisational risk appetite and accountability for verifying that operational reality matches that stated appetite. Automated systems can surface a risk tradeoff; they cannot own the decision. For the next generation of security leaders in India’s technology and BFSI sectors, the highest-value skill to develop is fluency in translating technical risk into business and regulatory decisions. This capability remains constant even as the underlying toolset continues to change.
