Practical Experience of Implementing Agentic SOC Platforms Abroad

【Introduction】 On October 28, 2025, the newly revised “Cybersecurity Law” was promulgated. The newly added Article 20 clearly states that “the state supports innovative network security management methods, utilizing new technologies such as artificial intelligence to enhance the level of network security protection“. The latest AI-enabled security operations represented by Agentic AI align with the requirements of the “Cybersecurity Law”.Currently, the Agentic AI-enabled Agentic SOC (Autonomous SOC) platform (i.e., Agentic SOP) is developing rapidly, continuously reaching new heights of hype. The Agentic SOC platform usesLLM as the central thinking unit, possessing autonomous reasoning, planning, and decision-making capabilities, able to invoke various tools to automatically complete scheduled security operation tasks, and achieve normalized security operation goals through human-machine collaboration.Previously, I have outlined the evolution process of SOC from manual to automated to autonomous, classified the levels of SOC autonomy, detailed the five key technical features of Agentic SOP, introduced Gartner’s interpretation of Agentic AI, and analyzed the latest developments of major foreign Agentic SOC platform vendors.This article will share a discussion on the practical implementation of the Agentic SOC platform published on the CSO Online website【Note 1】. Through interviews with security executives, product leaders, and researchers involved in pilot and deployment projects enabled by Agentic AI, or providing consulting for related projects, we gain a clearer understanding of the current state of Agentic SOC. Additionally, this article includes my comments on the interview information.The interview content covers six topics.

Advantages and Limitations of Autonomous AI

Advantages

Unlike traditional tools that only flag behaviors for review, agent-based systems “can handle L1 tasks, such as alert classification, cross-tool signal correlation, and in some scenarios, even take threat containment measures (like isolating endpoints), allowing analysts to focus on more strategic and important work”. The core value of autonomous AI lies in liberating human analysts from “repetitive drudgery”, allowing them to focus on higher-level exploration and threat hunting tasks. —— Jonathan Garini, CEO and Enterprise AI Strategist at FifthElement

In a security operations center environment, AI agents operate “similarly to digital L1 analysts — filtering data, collecting contextual information, and even generating detailed reports about their own activities”. He also mentioned the practical applications of AI agents in malware detection, script de-obfuscation, and tool collaboration. Teams that introduce agents can achieve “faster response times, a more streamlined team structure, and greater resilience in processing massive alerts”. —— Vinod Goje, AI Expert in Data-Driven Solutions and Applications at Bank of America

Autonomous AI excels at “responding to urgent scenarios in the ‘first 15 minutes’ — gathering contextual information, verifying threat intelligence, summarizing logs, and proposing actions for review”. Additionally, it can assist in vulnerability exposure management through prioritization and handle basic security maintenance tasks such as identifying inactive accounts. —— Itay Glick, Vice President of Products at OPSWAT

AI agents effectively mitigate “alert fatigue” by aggregating alert patterns and correlating them with threat intelligence sources; at the same time, NLP-based tools can achieve large-scale alert aggregation. —— Dipto Chakravarty, Chief Product and Technology Officer at Black Duck

Limitations

If there is a lack of clean data or clear scripts (workflows), agents may fall into ineffective information interference, or even autonomously “create” process steps —— Itay Glick, Vice President of Products at OPSWAT

Agentic AI has issues with false positives and overfitting.—— Dipto Chakravarty, Chief Product and Technology Officer at Black Duck

Even the most well-trained agents may struggle with ambiguous signals or multi-layered contexts. —— Prashant Jagwani, Senior Vice President and Global Head of Cybersecurity Services at Mphasis

From the experts’ statements, it can be seen that although Agentic AI has numerous advantages, its current practical application mainly shows significant effects on the alert assessment process of L1 analysts. The current level of practical application indicates that most enterprises still view Agentic AI as an auxiliary tool for human analysts rather than a replacement.

As Gartner stated, the value of generative AI in security operations still heavily relies on the prior skill level of SOC analysts. Either, during pre-training and fine-tuning, the LLM is infused with a large amount of high-quality prior security operation knowledge to obtain the so-called security domain model (which Gartner refers to as DSLM), and applied to Agentic AI; or, when using LLM for reasoning, the agent is provided with rich prior security operation knowledge and combined with contextual engineering to make LLM think more like a seasoned analyst.

At the same time, current Agentic AI still has many limitations but is rapidly evolving. Gartner warns that by 2027, 30% of SOC leaders will be unable to integrate generative AI into production processes due to inaccurate outputs and hallucinations.

Integration deployment models for agents: Additive vs. Standalone (Overlay vs. Replacement)

• Additive Model: Treats agents as extensions of SOP, SIEM, SOAR, or other security tools, achieving quick results with minimal disruption.

• Standalone Model: Here, the agent is located in a separate orchestration layer, offering greater flexibility but requiring stricter governance, more complex integration work, and more comprehensive change management.

Ready-made additive models built directly on SIEM or SOAR platforms usually perform best, while standalone frameworks “often require more effort for orchestration and governance” —— Jonathan Garini, CEO and Enterprise AI Strategist at FifthElement

“Microsoft’s Security Copilot helps analysts automatically classify alerts, filter out invalid information; CrowdStrike is also doing similar work; Google has launched an agent based on Gemini that can complete end-to-end alert investigations. Currently, the industry mainstream is still in the ‘additive and extension’ phase.” One reason for the popularity of additive components is that replacing or deeply integrating a new security operation platform is a daunting task — “deployment, personnel retraining, and process adjustments may take months, during which the team still needs to respond to real-time threats.” —— Amit Weigman, Cybersecurity and AI Expert at Checkpoint

The choice between the two models is a trade-off between “speed and flexibility”. “Additive is easier to adopt quickly, but less dynamic; standalone systems have stronger control but require more deployment and maintenance work.” —— Fergal Glynn, Chief Marketing Officer and AI Security Advocate at Mindgard

Proposed a “rule of thumb”: if most data exists within the current SIEM/SOAR processes, the additive model is the better choice; if data is scattered across IT, operational technology (OT), cloud, and software as a service (SaaS), the standalone agent layer is more effective, and “can reduce the hassle of ‘frequent system switching'”. —— Itay Glick, Vice President of Products at OPSWAT

Most enterprises start with the additive model — leveraging existing investments while testing effectiveness in a controlled environment. The standalone framework is usually a subsequent step, deployed only when enterprises are ready for centralized management across hybrid or multi-cloud environments. “From customer projects, we summarize an experience that many SOCs underestimate the complexity of integration. This is not only an issue of API connection systems but also requires aligning the decision logic of the agent with existing scripts and risk tolerance. The additive component model provides a smoother path for this alignment, while standalone orchestration is often a second-phase task after maturity improvement.” —— Prashant Jagwani, Senior Vice President and Global Head of Cybersecurity Services at Mphasis

The SOC platform is a complex, iterative evolution platform. For users of existing SOC platforms, how to gradually evolve to the new Agentic SOP stage requires careful path planning based on their own goals and specific circumstances. Therefore, the current Agentic SOP can be divided into two deployment models: overlay deployment and replacement deployment.

Overlay deployment is equivalent to the aforementioned “additive model”, which means deploying an overlay Agentic SOP on the existing SOP, obtaining alerts, events, and other information from the existing SOP, combining it with available contextual information, and using Agentic AI for analysis, assessment, investigation, response, and reporting, then returning the results to the existing SOP. At this time, the user’s security operation organization and processes remain basically unchanged, and Agentic AI empowers the existing SOP.Overlay Agentic SOP can be considered a trimmed version of SOP, focusing on the autonomy of security operation tasks, with a technical architecture closer to an agent management and operation platform.

Overlay Agentic SOP can be divided into two categories. One category is specialized products, often referred to as “AI SOC Analysts” abroad (Gartner calls them “AI SOC Agents”, i.e., “SOC Agents”), mainly used to automate and autonomize the work of L1 analysts (some also target L2/L3 analysts), often delivered in a SaaS model. The other category is a modular functional component (subsystem) that can be split from a complete Agentic SOP product.

Replacement deployment is equivalent to the aforementioned “standalone model”. Replacement Agentic SOP has a more complete architecture and functionality, with capabilities for cross-domain multi-source data collection, aggregation, storage, analysis, etc., possessing a relatively complete security data center and comprehensive native AI security operation functions, and embedding multi-agent management and operation subsystems within the platform. When adopting replacement deployment, Agentic SOP usually completely or partially replaces the existing SOP. At this time, the user’s security operation platform faces significant upgrades, which may involve organizational structure and process adjustments, data migration, etc.

Both models have their pros and cons. If users are relatively satisfied with their existing SOP, they can prioritize adopting the overlay Agentic SOP; if they are dissatisfied with certain parts (or even all) of the existing SOP, they can consider adopting the replacement Agentic SOP. Of course, if users are to establish a new SOP, they should prioritize considering the replacement Agentic SOP.

Currently, many users both domestically and abroad are dissatisfied with the security data center part of their existing SOP, which is also the root cause of alert fatigue. This issue cannot be resolved solely by Agentic AI (loading AI on low-quality data still yields unsatisfactory results); adjustments to the technical architecture of the security data center are needed. In other words, if the existing SOP data architecture is not adjusted, the empowering effect of Agentic AI is limited. Currently, mainstream Agentic SOPs have adopted a new generation of data-driven architecture. Therefore, more and more users are prioritizing the adoption of replacement Agentic SOPs to optimize the existing security operation data center architecture and expand Agentic AI capabilities on this basis.

Governance and Organizational Change

Practice shows that users prefer to gradually add new functions without affecting existing operations, so pilot projects have become a common first step.

The implementation of autonomous AI rarely happens overnight, “most security teams will not completely replace their existing security operations center with a brand new AI system — this is not only costly but also requires a lot of time and manpower investment, which may ultimately be difficult to advance due to excessive disruption and high costs.” —— Amit Weigman, Cybersecurity and AI Expert at Checkpoint

“For organizations intending to deploy autonomous AI, my primary advice is to start with pilot projects for small use cases, such as phishing attack response or credential abuse detection, and then gradually expand to broader detection and response scenarios.” Focusing on specific scenarios helps teams validate the technical value and reliability before large-scale changes. —— Jonathan Garini, CEO and Enterprise AI Strategist at FifthElement

Once agents are put into use, the governance system must also be upgraded accordingly. Teams will not discard existing frameworks but will adapt them: “Integrating existing change control and separation of duties rules into agent workflows — for example, destructive operations require dual signatures for confirmation, and determining ‘automatic execution’, ‘consult before execution’, or ‘escalate handling’ based on risk levels, and testing in a sandbox environment before formal deployment.” Additionally, agents are now included in red team testing, verifying security through prompt injection, “jailbreak” attempts, and other methods. “The core framework remains unchanged, but becomes clearer in the agent scenario.” —— Itay Glick, Vice President of Products at OPSWAT

Governance and risk control are expanded through “Human-in-the-Loop” approvals, rather than a complete reconstruction. Unless AI reaches a higher level of general intelligence, it is unrealistic to completely replace existing regulatory frameworks. —— Prashant Jagwani, Senior Vice President and Global Head of Cybersecurity Services at Mphasis

The application of Agentic AI in SecOps is not a simple application of tools or functions, but a systemic application, inevitably involving changes in the security operation system and the corresponding operational organizational changes. As I mentioned in the section 5.4 of the article “Towards the AI-Enabled SOC 4.0 Era”, the future of security operations must prioritize automation and intelligence, which will inevitably lead to a re-adaptation of the collaborative relationship between humans and machines in the security operation process, thus requiring us to reshape the governance system, organizational structure, and process specifications of operations.

In the article “Trends in Security Operations from Gartner’s 2025 North America Security Summit”, I introduced a diagram of Gartner’s analyst operational work migration, which also reflects the changes in operational organizations.

Some old positions have disappeared or weakened, while some new responsibilities have emerged. Of course, this process is not instantaneous but should be gradual, deepening continuously as the application of Agentic AI progresses.

Trust, Oversight, and Human-Machine Collaboration

The core advantage of autonomous AI lies in its autonomy, but this also becomes its main obstacle to implementation: many organizations are reluctant to let agents operate autonomously in production environments.

“If agents are misused or improperly deployed in threat response, a series of operations they execute may introduce new risks. For example, there may be unregulated script executions or exposure of newly discovered vulnerabilities.” Therefore, most organizations will not allow agents to operate completely autonomously without strong safeguards. It is recommended to view agents as “collaborative digital partners”, enhancing AI transparency through human-machine collaboration. —— Vinod Goje, AI Expert in Data-Driven Solutions and Applications at Bank of America

“AI still resembles a ‘black box’,” he said, “human analysts can also make mistakes, but managers are aware of the scope of their errors and can quantify the associated losses; with AI, we often ‘don’t know what we don’t know’, and this uncertainty naturally raises concerns.” Most users will “insist on human-machine collaboration in high-risk operations — AI can provide suggestions or classification information, but the final decision-making authority still belongs to the analyst”. Deploying specialized “narrow agents” focused on specific areas helps enhance visibility. “There is no need to build a large ‘black box AI brain’, but rather to create a set of specialized agents — each with a clear scope of responsibility, making monitoring and explanation easier.” —— Amit Weigman, Cybersecurity and AI Expert at Checkpoint

“All operations must have audit trails — from prompt inputs, tool calls, to output results and approval processes, without exception.” —— Itay Glick, Vice President of Products at OPSWAT

Documentation is crucial: “All regulated operations must be recorded, verified, and ultimately auditable. We must not only know what AI ‘did’, but also clarify ‘why it took these specific actions’.” —— Kyle Kurdziolek, Vice President of Security at BigID

Regulatory agencies in the financial services sector particularly emphasize “explainability in auditability”. “This means that AI outputs cannot merely be ‘black box suggestions’. Currently, teams are beginning to deploy layered audit tracking systems that break down the agent’s decisions into input data, confidence scores, and escalation logic.” —— Prashant Jagwani, Senior Vice President and Global Head of Cybersecurity Services at Mphasis

The article argues that the current issue with Agentic AI is its insufficient “transparency”, necessitating the establishment of “visibility” and “accountability” in AI workflows to enhance “transparency”. Meanwhile, having humans in the operational loop (Human in the operations loop, i.e., “human-machine collaboration”) is also crucial, as it reflects the current inadequacy of AI capabilities and is a necessity for maintaining human dominance in the future. Additionally, I believe that mechanisms and contingency plans for AI failures need to be established, ensuring that critical operational tasks can continue to function in the event of AI failures.

Clearly, to establish AI’s “transparency”, additional investments in AI are required. When considering investments in Agentic SOP, one should also account for the ancillary investments needed to establish “visibility” and “accountability” around Agentic AI. This cannot rely solely on the security and reliability (Security & Safety) mechanisms of Agentic SOP itself, nor solely on third-party AI security and reliability products or solutions, but also requires adjustments to the organization’s governance structure. This is one of the significant obstacles currently hindering the implementation of Agentic AI on the client side. In this sense, establishing Agentic SOP involves considerable investment, and the security and reliability investments of Agentic SOP can be coordinated with the overall AI investments of the enterprise.

New Generation Talent Development

If agents take over the work of L1 analysts, how should new members of the security operations center grow?

Traditionally, the work of L1 analysts has been the “training ground” in the security field. Autonomous AI brings a paradox: while it liberates humans from repetitive classification tasks, it may also weaken new analysts’ “muscle memory” accumulated through handling alerts. However, in reality, many mechanical classification tasks (such as filtering obvious false positives, cleaning up duplicate alerts, escalating routine phishing cases) can only train analysts’ patience and do not lead to substantial skill enhancement. AI excels at handling such trivial tasks, while human analysts can focus on more complex challenges. This shift transforms the L1 position from a “drudgery” into a “guidance training ground”: new analysts do not need to drown in ineffective information but can learn knowledge by studying cases organized and documented by AI and exploring the decision logic of agents. Therefore, without proper planning, autonomous AI may indeed lead to a talent development gap; but if utilized wisely, it can accelerate skill enhancement. —— Vinod Goje, AI Expert in Data-Driven Solutions and Applications at Bank of America

I believe Vinod Goje is relatively optimistic, while Gartner predicts“ by 2030, due to over-reliance on automation and AI, 75% of SOC teams’ foundational security analysis skills will degrade”【Note: This entry is included in Gartner’s eight major cybersecurity predictions for 2025】.How to avoid degradation? This requires rethinking the approach to talent development. As Vinod Goje stated, the growth path of junior security analysts has changed in the AI era, requiring them to maintain learning abilities, learn from AI, and learn from the results and processes of explainable AI assessments. Furthermore, it is essential to clarify which operational tasks should remain human-led, which require human-machine collaboration (HITL), and how to divide labor between humans and machines.Pricing, Value, and Solution Design

The capabilities and governance of autonomous AI are undoubtedly important, but one of the core drivers for its implementation in the security field is still “economics”. The primary concern of security leaders is: how much time and cost can it save us? The answer is often not simple.

The article suggests that this question may determine the future of autonomous AI in cybersecurity. Although the technology is rapidly maturing, its long-term viability ultimately depends on whether enterprises view it as a “sustainable solution for reshaping the operational model of security operations centers”.

“Pricing remains a friction point,” “suppliers are trying to adopt usage-based pricing models, but enterprises prefer to link spending to ‘savings in analyst hours’ rather than ‘raw computing resources or API call volumes’.” —— Jonathan Garini, CEO and Enterprise AI Strategist at FifthElement

Current AI pricing models vary significantly: “charging methods may include subscription, per seat, per alert, and some suppliers also offer usage-based plans. Advanced agent systems are usually priced higher due to their broader impact and ability to save analysts more workload.” —— Fergal Glynn, Chief Marketing Officer and AI Security Advocate at Mindgard

Some teams are experimenting with “per seat, per task, or hybrid models” for pricing, but: “hidden costs for storage, API fees, long prompt processing, script maintenance, etc., can accumulate quickly.” “Return on investment (ROI) should ultimately be reflected in specific metrics, such as faster detection and response times, more cases handled per analyst, and fewer false alerts.” —— Itay Glick, Vice President of Products at OPSWAT

Teams are “exploring pricing models comprehensively”, with usage-based and hybrid models continuously evolving. Enterprise budgets must cover not only software costs but also the mixed infrastructure costs of running large models both on-premises and in the cloud. —— Dipto Chakravarty, Chief Product and Technology Officer at Black Duck

Simple pricing metrics often overlook critical issues: “hidden costs often arise in specific areas such as retraining data models, cleaning structured telemetry data pipelines, etc.” The best return on investment comes from viewing agents as “part of a long-term process restructuring”, rather than “just another plugin”. —— Prashant Jagwani, Senior Vice President and Global Head of Cybersecurity Services at Mphasis

Measuring return on investment does not have a “unified standard”. “Every organization is different”. “Some organizations evaluate from an efficiency perspective: how many true positive/false positive events did the agent identify? How many incident investigations did it initiate? Others consider from a resource perspective: how much alert classification time was saved? How many times did the agent’s output require re-evaluation? Did the actual time savings meet expectations?” The core question is quite simple: can agents save enough time in classification and investigation work to allow security teams to focus more on enhancing the overall security capabilities of the enterprise? —— Kyle Kurdziolek, Vice President of Security at BigID

Pricing for security operation platforms increasingly needs to be linked to value, but this is not an easy task. The specific value of Agentic SOP is not only determined by vendor claims but also relates to the specific environments and application levels of each user.When evaluating return on investment, on one hand, the investment calculation should cover the “parts below the waterline of the iceberg”, including not only the investment in the Agentic SOP software system but also the investment in LLM (hardware and computing power consumption), AI security and reliability, and the investment in supporting security operation personnel (including in-house and outsourced). On the other hand, when calculating returns (value), a metric system can be used for evaluation. The metric system is closely related to the goals of project construction and does not have a unified standard, and cannot simply adopt a “copy-paste” approach.SummaryHere, I would like to borrow the conclusion of another article【Note 2】 as the conclusion of this article.

Agentic AI fundamentally changes our understanding of security automation — shifting from a rigid “rule-driven system” to a “collaborative partner” with adaptability and reasoning capabilities. But like all powerful tools, its value entirely depends on the prudence and responsibility of its deployment.

Those enterprises that successfully applyAgentic AI in SOC often resist the temptation to view it as a panacea for all security issues. Instead, they see it as a “complex capability that requires careful integration, continuous oversight, and ongoing optimization”.

The future development path is not a choice between “human analysts and agents”, but rather “building human-machine hybrid teams” — allowing both to leverage their unique strengths. Agents bring speed, stability, and the ability to process vast amounts of information in parallel; human analysts contribute creativity, contextual understanding, ethical judgment, and the ability to handle “undefined” ambiguous scenarios.

The future of SOC operations lies in this collaboration — but the premise is that we can build the necessary governance, security, and monitoring frameworks to ensure its reliable and secure operation.Agentic AI is not a “silver bullet” for solving cybersecurity challenges, but if deployed correctly, it will become a powerful tool for building “more resilient, faster responding, and more efficient security operations”.

Today, the question is no longer whether “Agentic AI will change SOC operations” — it is already changing. The real question is: can enterprises adjust their AI deployment strategies quickly enough to seize opportunities while mitigating risks? Those enterprises that balance this well will gain a significant competitive advantage in the ongoing battle against cyber threats; while those that rush to embrace the future of security automation but overlook risks may inadvertently create new vulnerabilities for themselves.

【Note 1】 The interview information mentioned in the text is sourced from the article:Agentic AI in IT security: Where expectations meet reality, link:

https://www.csoonline.com/article/4064158/agentic-ai-in-it-security-where-expectations-meet-reality.html

【Note 2】 The concluding quote is from the blog article: The Agentic AI Revolution in SOC: Promise, Peril, and the Path Forward, link:

https://www.sisainfosec.com/blogs/the-agentic-ai-revolution-in-soc/

Recommended Reading

• Cybersecurity Think Tank Platform is Long-Term Hiring Part-Time Researchers

• Welcome to Join the “Security Reference Hot Discussion Group”

Source: Focus on Security Management Platform

Click the card below to follow us,and let us help you understand cybersecurity ↓