Agentic AI shifts SOC analyst work from triage to oversight

Agentic AI systems now assume the high-volume SOC analyst tasks that once defined the job — alert triage, threat-intelligence correlation, and standard decision-tree workflow execution — shifting analysts from manual processing toward orchestration and oversight. The shift shows up in commercial products like Driven Tech’s Driven ARMOR, an agentic cognitive security operations platform powered by Torq that correlates telemetry across identity, endpoint, network, cloud, applications, and security platforms before executing intelligent response actions. Driven ARMOR is part of Driven Tech’s Pervasive Security platform, and Driven was recently selected for Anthropic’s Cyber Verification Program. The platform uses a vendor-agnostic architecture, integrates with Cisco, Palo Alto Networks, Microsoft, CrowdStrike, Splunk, SentinelOne, Rubrik, Google/Wiz, and other cybersecurity platforms, and its makers say it reduces alert fatigue, speeds up incident response, and strengthens cyber resilience without requiring organizations to replace existing security tools.

Agentic security operations retains experienced security professionals for critical decisions, while people validate outputs and serve as the final check on model results. Human oversight remains central in agentic security operations. Marci Paino, chief learning officer at Cisco and leader of the AI Workforce Consortium, said: “People remain the final check, validating outputs, assessing risk, and ensuring decisions are sound.”

The AI Workforce Consortium says 78% of cybersecurity jobs require AI skills, making AI literacy a baseline hiring expectation. The consortium found that from October 2025 to March 2026, AI skills appeared in 28.5% of G7 cybersecurity job postings, double the 14.2% share from the same period a year earlier. It also identified that SOC analysts now need expertise in overseeing AI agents and validating their outputs, while core human skills replace SIEM (security information and event management) platform skills as the baseline expectation in the agentic era. The consortium advises shifting hiring practices from experience-based models to skills-focused, proactive talent development programs.

The consortium identified new SOC analyst skills including AI-layer telemetry investigation, AI agent supervision, AI governance and compliance, AI-assisted threat hunting, and risk-based decision-making, representing the specific new capabilities expected of SOC analysts in the agentic era. It also found that analysts must develop the ability to validate AI outputs, orchestrate security workflows, and detect and respond to prompt injection attacks. The consortium identifies an emerging AI forensics role that investigates AI systems after failures, harm, or legal challenges, marking a new cybersecurity specialty created by the spread of agentic AI. The consortium found critical workforce gaps in generative AI, large language models, prompt engineering, AI ethics, and AI security.

Cited development programs include Google’s Vantage Project, IBM’s AI Builders Challenge, Cisco’s Networking Academy Capture the Flag initiative, and Pearson’s AI Skills for the Future.

AI forensics evidence differs sharply from traditional digital evidence. Instead of disk images or packet captures, investigators examine training data, model weights, prompt histories, and inference logs, reading them for bias, poisoning, tampering, and adversarial manipulation. The EU AI Act requires high-risk AI systems to maintain automated event logs throughout their operational lifespan to enable incident reconstruction. The consortium says the role demands expertise in digital forensics, AI platforms, applied AI, Python, SQL, data science, incident response, threat intelligence, machine learning, e-discovery, legal analysis, and compliance.

A Cisco survey found more than a third of cybersecurity leaders will invest in AI-powered security tools within one to two years, while only 25% prioritize spending on people, highlighting an imbalance between technology investment and workforce investment. The consortium found that human skills are growing rapidly in demand: ethical reasoning is up 533% year over year, systems thinking up 251%, and stakeholder engagement up 125%, reflecting growth in demand for non-technical human skills.

Security teams are reorganizing to integrate human and AI collaboration, adopting flatter structures and flexible reporting, with workers managing multiple AI agents, the consortium found. Cisco’s State of Wireless 2026 report found that 86% of organizations have difficulty hiring qualified wireless professionals, while 50% of respondents said AI is the top field pulling professionals away from wireless.

An IBM and Ponemon Institute study found that 85% of organizations boosted security spending due to frontier AI capabilities, more than the 64% that acted only after a breach. An AlphaSense survey of 1,028 leaders and AI decision-makers in 11 countries found enterprise buyers rank data privacy and security second only to accuracy when evaluating AI tools. AlphaSense expanded its enterprise AI security and governance program around its SuperAnalyst agent, adding hardened infrastructure, edge protection, pipeline guardrails, zero data retention, and certifications including SOC 2 Type II and ISO/IEC 27001. Pieter VanIperen, Chief Security and Information Officer at AlphaSense, said: “businesses should be able to adopt the latest AI innovations without needing to compromise on security or trust.”

F5 introduced Workforce AI Security to oversee and regulate both workforce and agentic AI usage across enterprise environments. F5 also upgraded Distributed Cloud Bot Defense to include persistent device intelligence and protections tailored for AI agents.

A Microsoft study found that 78% of users bring personal AI tools to the workplace, more often at small and midsize companies, reflecting the prevalence of unsanctioned AI tool use at work. A National Cybersecurity Alliance and CybSafe survey found that 43% of workers using AI on the job share sensitive information with AI applications without telling their employer.

Naynesh Patel, managing director of cybersecurity at Accenture, said data security failures typically stem from identity and governance problems operating at machine speed instead of from model failures. Patel recommends restricting external posting of shared information, keeping data inputs inside organizational boundaries, granting just-in-time least-privileged access, and using telemetry so SecOps (the security operations team) can see prompts and intervene at the moment of risk.

In 2023, Samsung suffered AI-related security incidents after an engineer pasted proprietary semiconductor source code into ChatGPT and another employee fed high-level meeting content into the tool. After those incidents, Samsung pursued disciplinary action, built an internal AI system with data controls, and enhanced security protocols.

Google announced Gemini 4 Carbon on October 29, 2026, a frontier model the company says leads rivals on benchmarks for detecting, exploiting, and patching vulnerabilities. Google is opening Carbon access to a select group of cyber defenders, with wider availability for approved partners through the Fairwind Program, which includes more than 650 organizations. Carbon follows Gemini 4 Argon, launched two weeks earlier; some Fairwind partners already have exclusive access to Argon’s cybersecurity defense capabilities.

Google reports Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on Collinear’s CWE-bench patching benchmark, topping OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in Google’s comparison. The company states it obtained these benchmark results with no guardrails in place. The company also states that cybersecurity teams now use AI models to assess which vulnerabilities are actively exploitable before prioritizing remediation efforts.

Wiz, a Google-owned cloud security firm, saw Carbon rank first on web penetration testing benchmarks. Carbon generated about half the unsupported claims of Gemini 4 Argon, Google reports, with internal teams observing an 80% reduction in high-severity misalignment alerts. The company reports that Gray Swan’s indirect prompt injection benchmark produced no successful attacks against Carbon, while Argon showed a 0.7% success rate. The company says Carbon’s cheating attempts on Terminal-Bench 4.0 fell to 0.4%, and none of them succeeded.

Google is participating in the US government’s voluntary pre-release model testing effort and will iterate on guardrails before wider developer, business, and consumer availability. Google will price Carbon at Argon’s introductory rates of $2 per million input tokens and $10 per million output tokens during the initial period, then move to $4 and $20 afterward.