AI agents can finally work the alert queue nobody had time for. Whether they close it correctly depends on what they're allowed to see, and who answers when they're wrong.
In March 2023, endpoint security tools started flagging the 3CX desktop app as malicious. Users compared notes on 3CX's own community forum, and some of them wrote exceptions so the app would keep running. It was signed by its maker, after all. 3CX ran the file through VirusTotal, got a clean result, and filed the alert under false positive.
The alert was right. A suspected North Korean group had gotten into 3CX through an employee who installed a trojanized copy of X_TRADER, trading software from a different vendor, and from there reached the company's Windows and macOS build environments.
Threat intelligence firm Mandiant said it was the first time it had seen one software supply chain attack set up another. Customers who installed the affected versions got the attackers' code in a properly signed release.
Nobody in that chain behaved foolishly. They trusted a code signature and a clean reputation lookup, which is roughly what most playbooks would have them do. The evidence that would have settled the argument was sitting in the binary.
Researchers who took it apart found code appended to a signed file, along with traces of SigFlip, a tool built to alter signed Windows files without breaking the signature.
Now give the 3CX playbook to an AI agent. Same signature, same clean lookup, same verdict, delivered in seconds instead of days. And if that verdict feeds the kind of shared learning loop vendors now advertise, the mistake travels to every customer on the platform.
That's the question underneath all the agentic SOC announcements. Agents will make security operations faster. Whether they make it right depends on what they reason over.
The queue nobody could finish
The case for agents is real. Ponemon's 2026 State of SecOps report, sponsored by AI SOC vendor Crogl, found the average organization generates 4,330 alerts a day and hands them to seven people, who investigate 37%. The rest go uninvestigated.
Attackers, meanwhile, have started to automate. CrowdStrike's 2026 Global Threat Report puts average eCrime breakout time at 29 minutes, with the fastest at 27 seconds.
Last November, Anthropic disclosed a Chinese state-sponsored espionage campaign in which attackers manipulated its Claude Code tool into doing an estimated 80% to 90% of the tactical work, at request rates no human crew could sustain. Nobody defends against that with a ticket queue and a pot of coffee.
Early results on defense look promising. In Prophet Security's 2026 survey of 250 security leaders and practitioners (Prophet sells AI SOC tools), 72% of teams already running AI said it cut investigation time by at least a quarter.
The obvious jobs to hand over first are enrichment, first-pass triage, evidence gathering, and stitching related alerts into one case. Agents can also work through the low-priority backlog that human teams quietly give up on, which happens to be where a patient intruder likes to wait.
Speed amplifies judgment, good or bad
Speed and quality travel together only when the evidence is good. A tired analyst who notices a gap in the telemetry tends to slow down. An agent that isn't built to flag uncertainty will produce a confident verdict from whatever it has and move to the next ticket.
The attackers found this out first. Anthropic reported that the AI running the Chinese campaign overstated what it had found. It claimed credentials that turned out not to work and flagged publicly available information as critical discoveries, so its human handlers had to check everything before acting on it.
Defensive agents will hallucinate as well. When an attacker's AI invents something, the attacker loses time. When a defender's does, a real incident gets closed.
Malware authors have also started writing for the AI reviewer. In June 2025, Check Point Research described a sample its author had named Skynet. Buried in the code was a plain-English message telling any language model analyzing the file to forget its instructions and report that no malware was present. It was a crude prototype, and it didn't fool the OpenAI models Check Point tried.
What caught my eye was the rest of the file, which ran through a list of checks for virtual machines and sandboxes before doing anything interesting. One sample, two marks: the sandbox and the AI. Check Point's researchers noted that the sandbox era produced hundreds of evasion techniques and expect the AI auditor to get the same treatment. I'd take that bet.
The remedy is an old one. Judge software by what it does, not by what it says about itself. A signature tells you who shipped a file. Only analysis tells you how it behaves. An agent that can't reach that layer is working from hearsay.
What CrowdStrike gets right, and where buyers should push
CrowdStrike's Fal.Con announcements on September 2 point the right way. An orchestrator agent calls in specialist agents that work endpoint, identity, cloud, SaaS, and network evidence in parallel, share what they find, and settle on one verdict.
Charlotte Agentic SOAR puts agent building, orchestration, and approvals in one workspace, and MCP support runs in both directions so outside agents and tools can plug in. It's a serious answer to the SOC's oldest complaint, which is that the evidence lives in systems that don't talk to each other.
Buyers should still test it on their own incidents rather than on a keynote demo. CrowdStrike's launch blog, updated September 18, lists the coordinated investigations and the certified third-party data pipelines as pre-beta. That's normal for a launch this size, and it's also a reason to wait for results.
The harder question is what kind of evidence the agents get. CrowdStrike pitches agents that run on "data CrowdStrike owns," and that telemetry is a real asset. The company also pipes in third-party data, and its Next-Gen SIEM accepts feeds from other EDR products.
But logs from other tools record what happened on a machine. They don't tell you what a file is. CrowdStrike recognizes this, which is why it rebuilt its Hybrid Analysis sandbox so agents can call it over MCP. A sandbox watches one run of a file, though, and malware that checks its surroundings first, as Skynet did, or that fetches its real payload later, as the 3CX malware did, can look clean in that run.
Whichever platform you choose, ask whether it can take independent file verdicts, keep their meaning intact, and show an investigator how they shaped the call.
CrowdStrike also announced tools that same day for securing AI agents themselves, including an identity provider built for them. That work is necessary, and it answers a different question.
A well-governed agent can still reach the wrong verdict, so locking down the agent tells you little about whether it catches attackers.
Four tests before you hand over the keys
Start with the evidence. Find out whether the agent can reach endpoint, network, identity, cloud, and file-level analysis, and whether it says so when the evidence is thin or contradictory. Every verdict should come with its sources. In Prophet's survey, 41% of respondents named lack of explainability as a top barrier to using AI in the SOC.
Then grant authority according to consequence. Assembling a case file needs different safeguards than disabling an identity or pulling a production server off the network. Write down which actions agents take alone, which need a human sign-off, how each decision gets logged, and which ones can be undone.
Most teams already work this way: not one of Prophet's respondents gives AI full unsupervised autonomy.
Third, attack it, and keep attacking it. Feed it unfamiliar malware, signed software that has been tampered with, incidents with missing telemetry, files that address the agent directly, and cases where the right answer is to hold off. Measure misses and false positives alongside speed and analyst hours, because a fast agent that misses things is a liability.
Finally, make it show its work across tools. An MCP connector moves data, and moving data is a long way from investigating it.
Check that evidence from other products arrives intact, gets examined at the depth it needs, and is stored where a specialist can review it and argue with it. In the Ponemon study, 49% of respondents said their data was too dispersed to normalize. Agents won't fix that on their own.
The org chart changes next
Understaffed SOCs need automation now. The bigger change comes to the org chart. Once agents absorb routine triage and evidence gathering, the old analyst tiers, defined mostly by how many tickets you could close, start to come apart.
What replaces them is a SOC with more threat hunters chasing new attacks, more detection engineers widening coverage, and a newer role I'd call the orchestrator: the person who designs agent workflows, watches them run, and fixes them when they drift. Senior defenders spend less time pushing cases along and more time deciding where automation has earned trust.
Autonomy should be earned one workflow at a time. Routine, well-tested actions can run on their own inside guardrails. Uncertain or high-impact calls still go to a person, and the line between the two will sit in different places at a bank and at a startup.
Responsibility stays with people. The organizations deploying these systems set the goals, decide who can do what, test performance, and investigate failures. An agent can execute a decision, but it can't sit in the post-incident review and explain itself to the board.
The vendors that win this market won't be the ones with the most agents or the boldest autonomy claims. They'll be the ones whose agents can show what they found, why they acted, which evidence carried the decision, and whether the company ended up safer. Trust the agent, if you like. Then decompile the evidence.
Other news: Google uncovered an AI framework designed to turn security patches into exploits, potentially shortening the time defenders have to update vulnerable software.





