Cisco's Security Strategy for Agentic Workforce: Closing the Trust Gap

In our first published video from ISMG Studio at RSAC 2026, Michael Novinson's discussion with Zscaler's Jay Chaudhry breaks down the complexities of building the infrastructure enterprises will need to harness the full potential of AI while staying within secured guardrails. Watch for Chaudhry's insights into containerization, inline inspection and the role for humans in taking the technology to scale. Stay tuned: More than 100 more interviews to come from ISMG before the conference leaves the Moscone Center. https://lnkd.in/e4T_hEjs #AI #Zscaler #ISMG #security #agenticAI #

AI agents are becoming a real security blind spot, and not enough organizations are treating them that way. Unlike human users, AI agents operate autonomously, often with persistent permissions that nobody is actively monitoring. If one gets compromised or misconfigured, the blast radius can be substantial, touching sensitive data, critical systems, or both. The challenge is that most deployments happen without centralized visibility or proper access controls. Teams move fast, agents get spun up, and governance catches up later, sometimes too late. The shift needed here is treating AI agents as non-human identities with the same rigor applied to privileged human accounts. That means just-in-time access, clear permission boundaries, and automated remediation when something looks off. Security and technology leaders who get ahead of this now will be in a much better position as agent adoption scales across the enterprise. #cybersecurity #AI

If “Claude Mythos” was exposed via a massive data leak, that’s not just an embarrassing pre-launch mishap - it’s a direct hit to the credibility of “safe, enterprise-ready AI.” When Anthropic is sitting at the top of the charts with Opus 4.6, the bar isn’t “can you build a strong model?” The bar is “can you run a secure operation while scaling?” This matters because AI labs are now critical suppliers. A leak can reveal capabilities to adversaries, accelerate misuse, and undermine trust with enterprise and government buyers. And it highlights the uncomfortable reality: alignment research doesn’t protect you from repo hygiene failures, insider risk, or sloppy access control. Operational security is AI safety. My controversial view: labs that can’t protect their own launch materials should not be trusted to deploy autonomous agents into enterprise workflows. If you can’t secure the docs, why should I believe you can secure the toolchain? Best practices: least-privilege access, secure-by-default repos, secrets scanning, red-team internal processes, incident drills, and public postmortems with corrective actions. Buyers should demand SOC2-style evidence and audit rights. If leaks become routine at frontier labs, will enterprises start treating model vendors like high-risk suppliers - or will they keep buying capability and hoping security catches up later? #Cybersecurity #AIGovernance #EnterpriseAI https://lnkd.in/gPxTTuVF

Most “AI-powered” security still relies on humans to glue everything together, stitching across alerts, tools, and domains. That’s not transformation. That’s AI as a helper. Arctic Wolf is pushing past that. In our latest blog, we introduce the Swarm of Experts™ — a coordinated network of specialized agents inside the Aurora® Agentic SOC, built not to replace analysts, but to reason, validate, and escalate the way real security teams do. Here’s what makes this approach different: 🧠 Coordinated specialization instead of generic autonomy 🕵️ Hypothesis-driven investigations instead of static playbooks 🔗 Shared context via the Security Operations Graph™ 🔄 Continuous operational feedback loops from 1,000+ security engineers The result? AI that doesn’t just respond faster; it thinks better. If you’re exploring the role of agents in the SOC and how to build trust in AI-driven investigations, this deep dive is essential. 👉 https://ow.ly/omkK50YCkip

What Happens When Attackers Feed Your AI SIEM a Fake Reality? Here’s a scenario that keeps me thinking as a detection engineer. An attacker compromises an environment. But instead of simply deleting logs to cover their tracks, they do something more dangerous: they generate synthetic telemetry. Fake authentication events. Fabricated network flows. Decoy process executions. Everything designed to look legitimate, but point investigators in the wrong direction. Now that data enters an AI-powered SIEM. Detection rules fire. Alerts trigger. The AI correlation engine does exactly what it was built to do — connect events, enrich context, assign severity, and build an investigation narrative. It looks convincing. It looks actionable. But it is built on false evidence. In a semi-autonomous SOC, where AI performs triage and investigation support but a human analyst makes the final call, this creates a serious risk. The analyst may receive a polished case summary with timelines, IOCs, ATT&CK mapping, and confidence scores — while the attacker’s real activity remains hidden behind synthetic noise. The AI has not “failed.” It has simply reasoned over poisoned inputs. Now take that one step further. In a fully autonomous SOC, the consequences could be even worse. An AI agent could isolate legitimate systems, block valid users, or trigger disruptive response actions based on fabricated telemetry. At that point, the attacker is not just evading detection — they are weaponizing the defender’s own automation. That raises a bigger question for all of us building or trusting AI in security operations: How do we establish trust in the data before we trust the AI’s conclusions? Should telemetry integrity become a first-class control in AI-driven SOCs? Do we need stronger provenance and chain-of-custody for logs? Can AI reliably detect when the very data it is analyzing was crafted to deceive it? Or are we still moving too fast toward autonomy in security operations? #CyberSecurity #SOC #SIEM #AISecurity #DetectionEngineering #BlueTeam #ThreatDetection #AdversarialAI #SecurityOperations

Why most AI security systems fail within hours under a real adversary AI security is built for stability. Attackers operate in instability. That gap is where systems collapse. Across 17 controlled adversarial simulations (SOC + AI-assisted pipelines): 68% showed successful lateral movement within the first hour Average response delay: 9–14 minutes In every case, detection occurred. Response failed. Failure didn’t start at detection. It started at control. One real scenario: Low-confidence anomaly flagged in EDR pipeline Analyst deprioritized due to alert fatigue Payload adapted within 12 minutes to evade signature update Privilege escalation → lateral movement → pre-escalation persistence No model failure. No tooling failure. Control failure. Attack loop: Detect → Delay → Adapt → Move → Escalate If this loop exists in your system, you are already compromised. You just haven’t detected it yet. This pattern repeats. I call it: Human Latency Collapse → response delay becomes attack surface Control Illusion Gap → system appears governed but breaks under pressure Adaptive Drift Exploit → attacker evolves faster than defense assumptions Most AI governance frameworks fail here. They optimize for compliance, explainability, reporting — not for real-time adversarial control. At VikSelora, systems are not secured first. They are broken first. AI models, SOC pipelines, and human workflows = one attack surface Simulations adapt in real time to defender behavior Failures are induced to map where control actually collapses Security is measured only after assumptions are removed VikSelora = adversarial AI systems lab focused on controllability under attack If your system has never been forced into contradiction, it is untested. Comment “SIM-01” + your domain (banking, healthcare, infra, defense, etc.). I’ll map a real failure scenario for it. Next: how a single analyst decision silently compromises an entire SOC chain. #AISecurity #AdversarialAI #CyberSecurity #SOC #AIGovernance #OffensiveSecurity #BlueTeam #RedTeam

As AI moves from experimental tech to critical infrastructure, the geostrategic value of intellectual property is rising sharply. At the same time, this shift is exposing gaps in both digital and physical infrastructure, with the private sector bearing much of the operational risk. CEOs already cite IP theft as a leading risk, and this threat is only set to intensify as AI accelerates the speed, scale and precision of cyber operations. Companies that adapt early and act decisively on data governance are in a better position to protect enterprise value and operate with confidence in an increasingly contested digital landscape.

I’ve been thinking about how quickly AI is becoming embedded in how governments operate and make decisions. It’s a defining moment for national security. Throughout my career, government technology has shifted from systems of record and back-office infrastructure to AI-led systems that can generate intelligence, surface risks, and inform action in real time. This Forbes piece highlights how AI is infiltrating and shaping the core of national security functions, including threat detection, intelligence analysis, and operational planning, while dramatically accelerating the speed at which insight can be generated and deployed. That shift raises important questions for me around trust, governance, accountability, and interoperability. When AI systems are influencing decisions at this level, it becomes critical to understand how they are designed, what data they rely on, and how their outputs are used. Just as important is whether they advance true interoperability, enabling coordination across agencies, partners, and coalitions so the full capability of the national security ecosystem can be brought to bear. Without that, even the most advanced technologies risk remaining fragmented and difficult to scale, rather than strengthening the collective effectiveness they are meant to support. Here is the full article: https://lnkd.in/eXqwHKsz #NationalSecurity #Cybersecurity #DefenseInnovation #Geopolitics #CyberDefense #NationalDefense #EmergingThreats #BoardofDirectors  #BoardGovernance #ArtificialIntelligence #AIEthics #ResponsibleAI

Most "AI-powered" security still relies on humans to glue everything together, stitching across alerts, tools, and domains. That's not transformation. That's AI as a helper. Arctic Wolf is pushing past that. In our latest blog, we introduce the Swarm of Experts™ — a coordinated network of specialized agents inside the Aurora® Agentic SOC, built not to replace analysts, but to reason, validate, and escalate the way real security teams do. Here's what makes this approach different: 🧠 Coordinated specialization instead of generic autonomy 🕵️ Hypothesis-driven investigations instead of static playbooks 🔗 Shared context via the Security Operations Graph™ 🔄 Continuous operational feedback loops from 1,000+ security engineers The result? AI that doesn't just respond faster; it thinks better. If you're exploring the role of agents in the SOC and how to build trust in AI-driven investigations, this deep dive is essential. 👉 Read: "Trustworthy AI Starts with Better Agents."

Most "AI-powered" security still relies on humans to glue everything together, stitching across alerts, tools, and domains. That's not transformation. That's AI as a helper. Arctic Wolf is pushing past that. In our latest blog, we introduce the Swarm of Experts™ — a coordinated network of specialized agents inside the Aurora® Agentic SOC, built not to replace analysts, but to reason, validate, and escalate the way real security teams do. Here's what makes this approach different: 🧠 Coordinated specialization instead of generic autonomy 🕵️ Hypothesis-driven investigations instead of static playbooks 🔗 Shared context via the Security Operations Graph™ 🔄 Continuous operational feedback loops from 1,000+ security engineers The result? AI that doesn't just respond faster; it thinks better. If you're exploring the role of agents in the SOC and how to build trust in AI-driven investigations, this deep dive is essential. 👉 Read: "Trustworthy AI Starts with Better Agents."