"Forty percent of security teams now use AI every day, and another 56% are testing it." That is Prophet Security's finding in the State of AI in Security Operations 2026 survey of more than 250 CISOs, SOC leaders and practitioners, and it frames a simple next question: if your team is ready to put AI agents into production, will your SOC make them effective — or merely noisier?
Which SOCs stand to gain most
Prophet Security maps the strongest use cases into two practical categories: organizations whose alerts are handled by an MDR or MSSP, and those whose own analysts carry the queue. For customers of an MDR or MSSP, the platform lists four common failures that an AI SOC agent can address: escalations that arrive hours after activity started; alerts returned for lack of context that require your analysts to finish the work; alerts closed without a transparent record of what was checked; and coverage that stops at the vendor's supported tools so your team must handle everything else.
For teams that work their own alerts — including what the provider returns — the survey enumerates a second set of operational strains: answering "are we impacted?" can take days; low- and medium-severity alerts wait or are closed in bulk; noisy detections get switched off; hunting and detection engineering only happen when someone has spare time; a few senior analysts carry the hardest cases; and nights and weekends are thinly staffed. The survey adds a stark statistic: 60% of respondents said an alert they had ignored or missed later turned into a serious issue such as a breach or downtime.
Where AI agents start on Prophet Security’s SOC Hierarchy of Needs
Prophet Security's SOC Hierarchy of Needs has five levels: alert management, detection coverage, threat awareness, proactive hunting and posture improvements. Each depends on the one below it. The company notes many SOCs have not stabilized even the first level; respondents said 28% of their alerts are never investigated.
AI SOC agents typically begin on the bottom layer — alert handling — and newer platforms also reach into proactive hunting and detection engineering. The vendor's point is procedural: agents free analyst hours at level one, and that time should be invested in the higher layers the SOC cannot currently reach.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleTechnical readiness: access, context and pipeline health
An AI agent only investigates what it can query. Prophet Security breaks technical readiness into three concrete requirements: API access to the SIEM, EDR, identity provider, cloud platforms, email and key SaaS applications; contextual data a system can read (asset criticality, user roles, VIP lists and known-benign activities); and healthy data pipelines so investigations do not miss silent or delayed logs.
The vendor suggests a practical test: list every query a senior analyst ran to close five recent incidents; any query the agent could not run marks a gap that must be closed before a pilot. The survey also records the most common roadblock: 44% of teams cited data privacy and concerns about how AI models are trained.
Define finished investigations, measure baselines, and reclaim detections
Before a pilot, Prophet Security recommends a written investigation standard for the highest-volume alert types — what evidence closes a phishing case, when impossible-travel warrants a manager call, and what suffices to call an alert benign. Teams should also inventory detections they have switched off; the survey found up to 40% of organizations had turned off some detections for capacity reasons.
Record baseline metrics you will compare against: the share of alerts investigated, mean time to investigate, time from alert to first look, and how often closed alerts are reopened. The report supplies reference points: a thorough investigation takes an average of 75 minutes, and alerts often wait close to an hour before anyone opens them.
Deciding agent permissions: approvals, widening authority, and oversight
Investigation and response carry different risks, and the survey shows how teams currently treat them. Fifty-seven percent require a human to review every AI decision before an alert is closed. For response, 44% have AI recommend actions that people carry out, 30% let it handle low-risk remediation on its own, and no respondent gives an AI full, unsupervised autonomy.
Prophet Security advises written rules up front: a list of actions that remain human-approved (for example isolating a host or disabling an account), a policy for who can widen the agent's permissions and under what track record, integration of findings into existing case management, and a named owner to feed agent mistakes back into its guidance. The recommended rollout mirrors human onboarding: start with read-only investigations, then widen permissions one category at a time.
What this means for MDRs, in-house SOC teams, and procurement
- MDRs and MSSPs: find escalation and context gaps that customers report — delayed alerts, missing context and opaque closures — and be prepared to demonstrate coverage and transparent audit trails if customers pilot agents.
- In-house SOC teams and analysts: prioritize API access and written investigation standards before a pilot; run agents in parallel and treat analyst overrides as training data about missing context.
- Procurement and security leaders: weigh build-versus-buy carefully. The survey notes 72% of AI users had tried building internal AI tools, and 46% of those projects were abandoned, never reached production, or were replaced by a commercial product.
Prophet Security packages these recommendations into products it calls Prophet AI SOC Analyst, Prophet AI Threat Hunter and Prophet AI Detection Engineer. According to the company, the SOC Analyst investigates every alert around the clock, queries SIEM, EDR, identity, cloud and email tools directly, and records every question, query and piece of evidence behind a determination. The Threat Hunter runs automated hunts against new severe threats and scheduled coverage-gap hunts. The Detection Engineer maps MITRE ATT&CK coverage from the organization's own investigations, proposes detections with backtests, and by default holds changes for team approval. Prophet Security also states customer data never trains its models and that the data plane can run inside the customer's VPC.
The practical next steps the vendor prescribes are concrete: close access gaps revealed by the five-incident test, write investigation standards, run the agent in parallel while recording baselines, and widen permissions only as the agent earns trust. For teams tempted to build their own agents, the survey's abandonment rate is a cautionary figure: prototype work is small compared with the ongoing engineering, maintenance and auditing that production requires.
