Proofpoint's SOC analyst agent investigates, never acts
Proofpoint's SOC Analyst Agent runs security investigations on OpenAI Daybreak models but takes no action itself. The restraint is the design lesson.
Proofpoint shipped an agentic AI security analyst into private preview this week, and the most interesting thing about it is a capability it deliberately does not have. The agent plans an investigation, pulls the evidence, and writes up a finding with a recommended next step. Then it stops. Disabling the account, blocking the user, containing the host — all of that still needs a person to press the button.
What actually happened
Per Proofpoint's announcement, the SOC Analyst Agent was built by the company's own Applied Services team and runs on OpenAI Daybreak models through the Daybreak Defense Network. An analyst describes a task in plain language — no query syntax, no jumping between four consoles. The agent plans the investigation, reads across connected Proofpoint data (alerts, logs, DLP events, user risk signals), and returns a structured finding.
Private preview covers data security, email DLP, insider threat management, and cloud DLP, with the rest of the product surface following next quarter. General availability is targeted for the end of Q3 2026. It also supports recurring investigations with routing and notifications, so the same question can run on a schedule.
Proofpoint is explicit about the boundary: the agent produces findings and recommendations. It does not take the action. OpenAI supplies the reasoning; Proofpoint keeps the data connections, workflows, and policy.
Why human-in-the-loop agent design matters for your business
You are not buying this product. You are, probably, building something shaped like it — an agent that reads your logs, your Stripe events, your support queue, your inventory table, and tells someone what is going on. And the temptation, once the reading works, is to let it write.
Resist that for one more quarter than feels necessary. The split Proofpoint drew is the correct default, and it is not about model quality. It is about blast radius. A wrong finding costs an analyst ten minutes. A wrong action — a disabled account, a cancelled order, a refunded invoice — costs you a customer and a postmortem. Read paths and write paths deserve different levels of proof, different credentials, and different review.
Practically: give the agent read-only tokens to start. Make every recommendation carry the evidence that produced it, so a human can check the reasoning in seconds rather than re-running the investigation. Log what it proposed alongside what the human decided. After a month, you will have a real dataset showing which recommendation types the human approves unchanged — and those, specifically, are the ones you can promote to automatic. That is how you earn write access. Not by trusting the model, but by measuring it.
Key takeaways
- Proofpoint's SOC Analyst Agent entered private preview on OpenAI Daybreak models, with GA targeted for the end of Q3 2026
- It plans investigations and returns structured findings across alerts, logs, DLP events, and user risk signals
- Containment, account changes, and every other action stay with a human by design
- Copy the split: read-only credentials first, write access only for recommendation types humans have approved unchanged
- Log the proposal next to the human decision — that log is what tells you when automation is safe
An agent that reads your systems is worth building. An agent that writes to them is worth building carefully. We design internal agents with the read and write paths separated, scoped credentials, and an audit trail you can actually read. See how we build agent workflows, or tell us what you want automated and we'll tell you which half to start with.
Sources: Proofpoint, Proofpoint press release.
- #ai-agents
- #security-operations
- #human-in-the-loop
- #agent-design
- #openai
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
XDOF at $1.2B: the training data is the product
XDOF is in talks at a $1.2B valuation for collecting robot training data, three months out of stealth. Proprietary training data is the asset labs cannot self-serve.
Read itHuskeys raised $27M because WAFs can't read AI agents
Blackstone led a $27M round for Huskeys, an AI layer over web application firewalls. AI agent traffic breaks bot detection built to answer one question: human or not?
Read it