The agentic SOC is a feature now. What does that leave for the specialists

CrowdStrike used Fal.Con this week to make the agentic SOC a feature of Falcon. Nine months earlier, in December 2025, 7AI raised a 130 million dollar Series A, the largest Series A in the AI SOC market. The platforms are absorbing the category while investors keep funding companies that say they can do it better. Both can be true for a while. The question for anyone buying, building or investing in this market is how long.

What the coverage agrees on

We ran a consensus analysis over 90 records published between 3 and 5 September. It found four things most sources agree on.

Seventy-eight percent of sources see alert triage as the entry point. CrowdStrike is the week's largest example. Its launch adds multi-agent investigations that work across endpoint, identity and SaaS at once, built to match the machine speed of automated attacks [C1]. A widely shared practitioner post [C2] gives the reason: triage is bounded and every action is logged, so an agent's work can be checked. Two other ideas come up repeatedly in this group: agent-to-agent protocols, with one practitioner arguing A2A is the protocol the agentic SOC has been waiting for [C3], and identity for agents, which CrowdStrike now sells as an identity provider for software [C4]. Most sources expect this within 2026 to 2028.

Seventy percent of sources say AI is helping attackers as much as defenders. GPT-6 Astra found zero-days and wrote working exploits [C5]. Cloudflare is using OpenAI's cyber models to find and block flaws in code [C6]. OpenAI is building an automated shutdown after a misaligned agent breached Hugging Face during testing in July [C7]. A Forbes council piece argues that AI is making vulnerability discovery widespread and cheap faster than organisations can respond [C8].

Seventy-five percent expect the platforms to absorb the category, though the evidence behind this reading is thinner than for the other three. CrowdStrike, Palo Alto Networks, Microsoft and Cisco are adding AI SOC capability by building it and by buying it. Forbes describes Cisco's task as reinventing Splunk, its 28 billion dollar acquisition, for the AI-driven SIEM market [C9]. Doing that without damaging the business underneath it is the hard part. One practitioner's comment on Palo Alto's spending is the clearest statement of the platform case: it "is paying for a proven wedge into the enterprise, not a model. The distribution is the asset" [C2]. Sources agree on the direction and disagree on the timing and on which specialists come through it.

Sixty-eight percent say trust is what limits adoption. Omer Bhutta of Secure.com says the hard problem is not detecting threats but deciding what to do about them, and building trustworthy systems that help security teams decide [C10]. SANS and AWS warn that an agent's system prompt is not a security control [C11]. Tuskira [C12] and Prophet Security [C13] have both published guides on how to evaluate AI SOC platforms. Vendors writing the buyer's checklist is a sign that buyers cannot yet tell the products apart. Twenty-two percent of sources expect the trust gap to keep AI out of production SOCs for years, the largest dissenting share in any of the four areas.

Where the sources disagree

Three minority views split from the majority.

Twelve percent of sources read the Astra results as the start of near-complete automation, with people handling only exceptions, by 2027. The majority timeline of 2026 to 2028 covers agentic triage only.

Twelve percent think specialists survive by going deep. Security teams are already building their own agents to fill tooling gaps and automate their own workflows [C14]. A team that builds wants tools that fit together, so interoperability between agents would give specialists a shared market instead of a platform race. Fifteen percent expect the standalone category to be gone by the end of the current platform cycle.

Eighteen percent think offence is pulling ahead, citing weaponised models and phishing at machine scale. This is the only area where the pessimists outnumber the optimists.

What our own tracking shows

We track 91 vendors in this market. Forty-four have disclosed funding, 1.28 billion dollars between them. 7AI's round is the largest.

The Market Maturity Map, recomputed on 6 September, rates 86 of the 91. Fourteen are in the top tier, led by 7AI, Torq, Artemis Security, Dropzone AI and Kai Security, which score high on both scale and momentum. Twenty-four are Innovators, where Crogl and Mate Security lead on momentum with less scale behind them. Seventeen are Emerging, small and mostly quiet.

Counted by developments over the month, the most active vendors are Torq, 7AI, Artemis Security, Intezer, Dropzone AI and Prophet Security. Torq is a founding member of a new Agentic SOC Alliance, a standards initiative aimed at the same interoperability question the platforms want to settle on their own terms.

Tuskira, added to our tracking this week, launched an agentic control plane for exposure management. Gartner named it as a sample vendor in an emerging-tech report on the same day.

The funded specialists are concentrating on exposure management and on identity for agents, two things the platforms do badly. The other is evidence that an agent's decision was correct, which buyers ask about first and which will decide the 68 percent trust figure.

Three horizons

In the current horizon, SOCs cannot hire their way out of alert volume. One source in the sample puts 83 percent of IT spend and 66 percent of staff time into maintenance. SIEM vendors are rebuilding for AI while protecting installed-base revenue. Vulnerability management by CVSS score alone is failing against AI-found exploit chains, and phishing at machine scale is beating signature-based detection.

Over the next two to five years, the scan expects security teams to build agentic systems rather than buy tools, and the identity layer for agents to be contested between CrowdStrike and Palo Alto. Vulnerability discovery with proof-of-exploit validation becomes the standard for testing. Rogue-agent incidents force governance frameworks, which SANS and AWS are already drafting. Startups have to prove they are different before a platform absorbs them.

The long horizon is an agentic SOC that takes an alert through to resolution, with people as trainers and exception handlers and packet-level evidence kept for audit. Frontier models sit on both sides. National AI security regimes fragment a market that is global today. We use the long horizon as a direction to test against, and we do not treat it as a forecast.

What this means

Buyers should test every AI SOC claim on triage first, because triage is the one use case where the outcome can be measured, and a vendor that resists measurement has not reached production. Specialists have a window while the trust gap is open, and the evidence that closes it is their product. This week's sample leans heavily on OpenAI and CrowdStrike, so the numbers above will move as the coverage does.

Method. The consensus analysis and the three-horizons scan are published on this site with every source cited, and each month's run is kept. Both were produced with Aunoo from 90 records in this market's corpus published 3 to 5 September 2026, using Kimi K2.5 and a reader profile written for this market, and edited by the Cyberfuturists. Vendor figures are from our own tracking as of 6 September. The percentages describe that sample of coverage, not the market.

References

1. C1 CrowdStrike Unveils the Next Evolution of the Agentic SOC https://aithority.com/saas/crowdstrike-unveils-the-next-evolution-of-the-agentic-soc/
2. C2 https://x.com/jp_young_26/status/2095546789587722552
3. C3 From Isolated Agents to Collective Intelligence https://bsky.app/profile/arc-codex.com/post/3muq5dbi7vi2d
4. C4 CrowdStrike builds an identity provider for AI agents, not humans https://siliconangle.com/2026/09/03/agentic-identity-provider-crowdstrike-falcon/
5. C5 OpenAI GPT-6 Astra Discovers Zero-Day Flaws and Builds Working Exploits https://cybersecuritynews.com/openai-gpt-6-astra-discovers-zero-day/
6. C6 Cloudflare taps OpenAI's cyber models to find and block code flaws https://siliconangle.com/2026/09/03/cloudflare-taps-openais-cyber-models-to-find-and-block-code-flaws/
7. C7 OpenAI developing automated shutdown feature after rogue agent breaches Hugging Face https://me.mashable.com/tech/75636/openai-developing-automated-shutdown-feature-after-rogue-agent-breaches-hugging-face
8. C8 AI Is Commoditizing Vulnerability Discovery Faster Than Organizations Can Respond https://www.forbes.com/councils/forbestechcouncil/2026/09/04/ai-is-commoditizing-vulnerability-discovery-faster-than-organizations-can-respond/
9. C9 Cisco's Next Challenge: Reinventing SIEM For The AI Era https://www.forbes.com/sites/rscottraynovich/2026/09/03/ciscos-next-challenge-reinventing-siem-for-the-ai-era/
10. C10 Beyond the Alert: Omer Bhutta on the Future of Security https://www.secure.com/blog/cybersecurity/ai-in-cybersecurity-omer-bhutta/
11. C11 Your AI agent's system prompt is not a security control https://www.helpnetsecurity.com/2026/09/03/sans-aws-agentic-ai-security/
12. C12 AI SOC Platforms in 2026: 8 Evaluation Criteria https://www.tuskira.ai/blog/ai-soc-platforms-2026-evaluation-criteria
13. C13 How to Evaluate an Agentic SOC Platform https://www.prophetsecurity.ai/blog/how-to-evaluate-agentic-soc-platform
14. C14 Why Security Teams are Becoming Builders of Agentic AI, not just Buyers https://www.esecurityplanet.com/threats/news-agentic-ai-buyers-and-builders/