Prophet Security

Prophet Security builds an agentic AI SOC — “a fleet of autonomous AI agents that accelerates Tier 1, Tier 2, and Tier 3 tasks, from alert investigation and response to proactive threat hunting” (home). Its agents “investigate alerts as an experienced human analyst would, using flexible, explainable reasoning, not rigid playbooks” (Series A). The engineering bet is reasoning over playbooks: for every alert, an LLM agent dynamically builds an investigation plan, pivots across the customer’s security stack to gather evidence, and decides true-vs-false-positive — then shows its work (plan, queries, evidence) so a human can trust it. Three products ride that engine: SOC Analyst, Threat Hunter, and Detection Advisor (Series A).

Vitals: Menlo Park + NYC (hybrid/remote) · $30M Series A (Accel) + $11M seed (Bain) + strategic (Amex, Citi) · ~50–80 people · SOC 2 Type 2.

Business context — founders, funding, customers, metrics

• Founders: Kamal Shah (CEO) and Vibhav Sreekanti (CTO); Grant Oviatt is VP Product (About, Series A signed “-Kamal & Vibhav”). Built “by SecOps experts” (home); founders “previously scal[ed]” security companies (About).
• $30M Series A led by Accel (Eric Wolford), with Bain Capital Ventures and other strategic investors (July 29 2025, Series A); a reported $11M seed (Bain Capital Ventures) preceded it. Later strategic investments from Amex Ventures and Citi Ventures (Feb 2026, Amex/Citi); named to the Redpoint InfraRed 100.
• Customers shown publicly: Instacart, Redis, Penske, Moveworks, Compass, Udemy, IAC, Thirty Madison, JB Poindexter, ETS, Partsource, Upgrade, Cabinetworks (home, Series A).
• Reported outcomes: “10x increase in SOC throughput,” “90% reduction in MTTI and MTTR,” “75% faster triage and investigation,” “100% Alert Coverage across all severities” (home).
• Positioning: a SOAR / MDR replacement — “reasoning-based investigation to eliminate playbook maintenance” and “role elevation, not role elimination” (Series A, blog).

The heavy lifting

• Reasoning replaces playbooks. For each alert the agent “dynamically builds an investigation plan, i.e. what are the key questions that an expert analyst would ask,” then executes it — so coverage doesn’t decay the way static SOAR playbooks do as upstream APIs and detections shift (How it works, blog).
• Investigation depth by pivoting across the whole stack. Agents “retriev[e], correlat[e], and analyz[e] all … context from multiple data sources (SIEMs, security data lakes, security tools, object storage)” and offer “Dig Deeper” follow-ups — depth, not just faster triage, is the stated accuracy lever (How it works).
• Show your work, then gate autonomy. “100% transparency” — every investigation exposes “the investigative plan, the queries used … and all the evidence gathered” — and remediation is autonomous for high-confidence threats, human-in-the-loop for complex cases (home).

Stack

A Python/Go backend, a React/TypeScript frontend, and an in-house agentic-AI platform. Rows are from the job board (R&D) and product pages; the LLM/cloud specifics aren’t named — see Likely internals.

Layer	Choice	Evidence
Backend languages	Python · Go	Backend JD, SecOps JD
Frontend	React · TypeScript	Full-Stack JD, Backend JD
AI platform	in-house Agentic AI platform (LLM agents: plan / investigate / respond)	ML JD, How it works
AI methods	prompt engineering, retrieval-based context augmentation, fine-tuning, safety	ML JD
Agent work	”AI-powered agents, data synthesis and correlation, security tool integration”	Backend JD
Data sources read	SIEMs, security data lakes, security tools, object storage	How it works
Integrations	bi-directional connectors: SIEM/EDR/IAM/SOAR/ITSM + case mgmt; named: Google Security Operations (Chronicle), ExtraHop (NDR)	home, blog
Outputs	severity + plain-English findings; remediation steps; dedup; write-back to case mgmt	How it works, Series A
Compliance	SOC 2 Type 2	home

Key finding — the ML team builds the platform, not a model

The Senior ML Engineer leads “the architecture, design, and development of our Agentic AI platform,” working in “prompt engineering, and retrieval-based context augmentation methods … fine-tuning techniques, and safety considerations” (ML JD). The differentiated asset is the reasoning/RAG platform over a customer’s security data — the foundation model is a (likely multi-vendor) dependency.

Hard problems

The parts an engineer at this company loses sleep over. Public signal is cited (verified); likely approach is labeled speculation — best-practice fill-in, hedged.

Problem	Why it’s hard	Public signal	Likely approach (speculative)
Investigation accuracy without playbooks	A missed true positive is a breach; reasoning over noisy, ambiguous alerts must beat both static playbooks and shallow triage	”flexible, explainable reasoning, not rigid playbooks”; “depth of investigation is the holy grail of AI SOC accuracy” (Series A, blog)	Plan-then-investigate agent that enumerates expert questions, gathers evidence per question, and abstains/escalates when inconclusive
Bi-directional integration across every SOC stack	Each customer runs a different SIEM/EDR/IAM/SOAR mix; ingestion alone isn’t enough — the agent must act and write back	connectors with “bi-directional communication to support the full investigation lifecycle, not just simple alert ingestion” (home); Google SecOps + ExtraHop named (blog)	A connector framework with read (query) + write (respond/case-update) per tool; normalize telemetry into a common investigation schema
Trust + gated autonomous remediation	Acting in a customer’s production security stack has real blast radius; over-trust causes harm, under-trust kills the value	”100% transparency” (plan + queries + evidence); autonomous for high-confidence, human-in-the-loop for complex (home)	Confidence-scored actions gated per blast radius; full reasoning trace persisted per investigation as the audit record
Evaluating a security agent + adapting per env	Quality across thousands of alert types can’t be unit-tested, and each environment’s “normal” differs	ML role owns “safety considerations”; the Adapt phase “learns from every analyst feedback” (ML JD, How it works)	Golden-incident eval sets + LLM-as-judge graded on analyst feedback; per-tenant context/policies injected via RAG to tune reasoning

Likely internals

The infrastructure Prophet doesn’t name publicly, inferred from the stack it does (Python/Go on a cloud, an LLM+RAG agent platform):

Component	Likely choice	Basis
LLM providers	OpenAI + Anthropic frontier models, routed	agentic reasoning + “fine-tuning” (ML JD); Prophet’s own blog analyzes Anthropic/Google model behavior; production vendor unnamed
RAG / context store	a vector index over org context, prior investigations, and playbooks	”retrieval-based context augmentation” + Adapt ingesting “organizational context” (ML JD, home)
Agent orchestration	an in-house planner/executor over a tool/connector layer	”Agentic AI platform” + plan→investigate→respond (ML JD, How it works); no named framework
Cloud	AWS	conventional for a Menlo Park R&D security SaaS; not stated
Primary DB	Postgres + an object/evidence store	investigations, evidence, case state; object storage is a named read source, write store unstated
Eval / observability	golden-incident eval harness + LLM-as-judge	”safety considerations” (ML JD); explainability is a core claim (blog)
Auth	enterprise SSO (SAML/OIDC), least-privilege read connectors	SOC 2 Type 2; POV uses “read-only access to 2-3 data sources” (How it works)

Architecture

The investigation loop: Plan → Investigate → Respond → Adapt

Prophet’s SOC Analyst runs a four-phase loop per alert (How it works). Plan: “instantly summarizes incoming alerts, extracts key artifacts, classifies them, and dynamically builds an investigation plan.” Investigate: “emulates an expert analyst, executing the investigation plan by retrieving, correlating, and analyzing all information … from multiple data sources (SIEMs, security data lakes, security tools, object storage),” with “Dig Deeper” for follow-up questions. Respond: “assigns severity … prioritizes critical alerts … delivers concrete remediation steps … deduplicates related alerts,” integrating with collaboration and case-management tools — autonomous for high-confidence threats, human-in-the-loop for complex ones (home). Adapt: “learns from every analyst feedback and continuously adapts to your environment.”

Mermaid source

flowchart LR
  classDef io fill:#fdf4e8,stroke:#d97706,stroke-width:1.5px,color:#0f172a;
  classDef ai fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a;
  classDef data fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a;
  classDef human fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a;

  Alert(["Alert<br/>SIEM · EDR · IAM · cloud · email"]):::io

  Plan("Plan<br/>summarize · extract artifacts · classify<br/>build investigation plan (expert questions)"):::ai
  Invest("Investigate<br/>execute plan: retrieve + correlate across<br/>SIEM · data lake · tools · object storage<br/>+ Dig Deeper"):::ai
  Decide{"Respond<br/>true vs false positive<br/>+ severity · confidence"}:::data

  Auto("Autonomous remediation<br/>high-confidence threats"):::ai
  Human("Human-in-the-loop<br/>complex / high-blast-radius"):::human
  Case[("Case mgmt + collaboration tools<br/>(bi-directional)")]:::data
  Adapt("Adapt<br/>learn from analyst feedback<br/>refine reasoning to env"):::ai

  Alert --> Plan --> Invest --> Decide
  Decide -->|high confidence| Auto --> Case
  Decide -->|complex| Human --> Case
  Case --> Adapt
  Adapt -. "tunes" .-> Plan
  Invest -. "shows plan + queries + evidence" .-> Case

The platform: three agents on one reasoning engine

The same reasoning engine powers three products: SOC Analyst (autonomous triage/investigation/response), Threat Hunter (natural-language and “always-on and scheduled” hunts over a “curated library of pre-codified hunt templates”), and Detection Advisor (“improve detection coverage and reduce alert noise based on actual telemetry”) (home, Series A). The engine reads and acts through bi-directional connectors into the customer’s SIEM/EDR/IAM/SOAR/ITSM stack — “built with bi-directional communication to support the full investigation lifecycle, not just simple alert ingestion” (home) — with Google Security Operations (Chronicle) and ExtraHop (NDR) among named integrations (blog).

Mermaid source

flowchart LR
  classDef sys fill:#eef2f8,stroke:#94a3b8,stroke-width:1.5px,color:#0f172a;
  classDef ai fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a;
  classDef data fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a;
  classDef human fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a;

  subgraph Stack["Customer security stack · bi-directional connectors"]
    direction TB
    SIEM("SIEM / data lake<br/>Google SecOps · Splunk-class"):::sys
    EDR("EDR · IAM · NDR (ExtraHop)"):::sys
    Case2("SOAR · ITSM · case mgmt · Slack"):::sys
  end

  subgraph Engine["Prophet AI · reasoning engine"]
    direction TB
    LLM("LLM agents<br/>plan · investigate · respond<br/>reasoning, not playbooks"):::ai
    RAG[("Org context + playbooks<br/>retrieval-augmented")]:::data
    LLM --- RAG
  end

  subgraph Agents["Agent products"]
    direction TB
    Analyst("SOC Analyst<br/>triage + investigate + respond"):::ai
    Hunter("Threat Hunter<br/>NL + scheduled hunts"):::ai
    Advisor("Detection Advisor<br/>tune detections from telemetry"):::ai
  end

  Evidence[("Transparency record<br/>plan · queries · evidence")]:::data
  Analystppl("Human analyst<br/>review + feedback"):::human

  Stack <--> Engine
  Engine --> Agents
  Agents --> Evidence
  Evidence --> Analystppl
  Analystppl -. "feedback → adapt" .-> Engine

Inference — explainability is the trust wedge against SOAR/MDR — confidence: high

Prophet’s whole pitch against incumbents is that the agent “show[s] their work” — Accel called it “a critical differentiator in an industry that runs on trust” (Series A). Persisting the plan, queries, and evidence per investigation isn’t a feature; it’s what lets a CISO let an agent act autonomously at all.

Team & process

A SecOps-expert-led R&D org — founders Kamal Shah (CEO) and Vibhav Sreekanti (CTO), with Grant Oviatt as VP Product — split between Menlo Park (HQ) and New York, hybrid/remote (About, Backend JD).

Role	Person	Source
Co-founder / CEO	Kamal Shah	About, Series A
Co-founder / CTO	Vibhav Sreekanti	About, Series A
VP Product	Grant Oviatt	About

Engineering is one R&D org (Backend, ML, Full-Stack, plus Core-Product and Platform engineering managers), comp $150–273K + equity, with a dedicated Security Operations Engineer who writes code “to support investigations or automation” — practitioners embedded in the build (Backend JD, ML JD, SecOps JD). The product itself is the human-AI process: Prophet frames the analyst’s new job as investigation reviewer — the agent does “the burdensome first steps of triage and investigation,” the human provides feedback during onboarding and per-investigation, and that feedback is the Adapt loop (Series A, How it works). Onboarding is a 30-minute POV on “read-only access to 2-3 data sources” (How it works) — a deliberately low-friction land-and-expand. A heavy practitioner-authored blog (SecOps veterans on staff) doubles as detection/threat content and recruiting signal (blog).

Sources

Reconstructed from public sources only — no insider information. Crawled 2026-06-09 via Chrome MCP (logged-out) + the Ashby posting API. First-party (prophetsecurity.ai, the Prophet blog, Prophet’s Ashby board) prioritized; Accel/press labeled third-party. Claim tiers: verified (stated on a public page, linked) · inferred (reasoned from a cited signal, confidence flagged) · speculative (best-practice fill-in, labeled). Links are live; pages change, so the supporting quote for each claim is kept in this repo’s evidence map (evidence/prophet-security-evidence-map.md).

#	Source	Link
S1	Homepage	https://www.prophetsecurity.ai/
S2	How it works (AI SOC Analyst)	https://www.prophetsecurity.ai/ai-soc-analyst
S3	About us	https://www.prophetsecurity.ai/about-us
S4	Blog index	https://www.prophetsecurity.ai/blog
S5	Blog — $30M Series A (Accel)	https://www.prophetsecurity.ai/blog/prophet-security-raises-30-million-series-a-led-by-accel
S6	Blog — Amex Ventures + Citi Ventures	https://www.prophetsecurity.ai/blog/accelerating-the-agentic-ai-soc-movement-with-amex-ventures-and-citi-ventures
S7	Job board (Ashby)	https://jobs.ashbyhq.com/prophet-security
S8	Senior Software Engineer, Backend (JD)	https://jobs.ashbyhq.com/prophet-security
S9	Senior Machine Learning Engineer (JD)	https://jobs.ashbyhq.com/prophet-security
S10	Software Engineer, Full Stack (JD)	https://jobs.ashbyhq.com/prophet-security
S11	Security Operations Engineer (JD)	https://jobs.ashbyhq.com/prophet-security