Doppel automates phishing takedowns with a five-stage GPT-5 and RFT pipeline

The economics Doppel is fighting

Doppel's premise starts with a timing problem. According to the company, a single impersonation site can launch, target thousands of users, and disappear in under an hour — and generative tools let attackers spin up hundreds of variants in seconds. The old model of digital risk protection, with humans manually reviewing suspect domains and social profiles, breaks when attacks arrive faster and across more surfaces than people can evaluate.

The ability to generate infinite persuasion at almost no cost changed everything.Montana Labs

That line from co-founder and CTO Rahul Madduluri frames the whole rollout. If attackers use AI to scale, a defense built on manual review loses by arithmetic. Doppel's response was to move the classification decision itself into models.

What the pipeline actually does at each stage

Doppel describes a five-stage flow rather than a single model call, and the division of labor is worth reading closely. Millions of domains, URLs, and accounts arrive daily. o4-mini plus heuristics filters noise and extracts structured features. Multiple GPT-5 prompts, each purpose-built for a threat type like impersonation risk or brand misuse, run in parallel to confirm intent. An RFT-tuned version of o4-mini then assigns a structured label — malicious, benign, or ambiguous. A second GPT-5 pass validates that decision and writes a natural-language justification, and if confidence clears a threshold, enforcement fires automatically.

The design choice here is that cheaper models handle filtering and final classification, while GPT-5 does the reasoning-heavy confirmation and verification. Low-confidence or conflicting cases route to human analysts, whose decisions feed back into training. It is a system engineered to spend expensive inference only where the ambiguity justifies it.

Why consistency, not accuracy, was the limiting factor

The most instructive detail in Doppel's account is what RFT was brought in to fix. The company had already seen gains from an LLM-enhanced pipeline. The wall it hit was that the same threat could be judged differently depending on which analyst looked at it. Reinforcement fine-tuning turned each analyst decision — malicious, benign, or unclear — into a graded example, training o4-mini to replicate expert judgment on edge cases.

One real benefit that came out of RFT is you're making that model's decisions more consistent.Montana Labs

Software engineer Kiran Arimilli's framing matters because consistency is often invisible in benchmark scores. Two analysts each being right most of the time can still produce a pipeline nobody can predict. Doppel went further by designing grader functions that rewarded explanatory quality, not just correct answers — models that reasoned clearly scored better than models that happened to guess right.

Auto-generated justifications as the trust mechanism

Doppel closes the loop by attaching an AI-generated justification to every automated takedown, explaining why a threat was removed. Previously that explanation required analyst intervention. This is the specific implication worth naming: in autonomous enforcement, the model output that unlocks the automation is not the classification itself but the explanation of it. Customers get the confidence to act quickly and the context to defend those decisions to stakeholders internally.

Doppel says domains are the hardest channel it handles — messy signals, constantly changing content, threats evolving across surfaces at once — and having largely automated it, the company plans to extend the same framework to social media and paid ads, scale its RFT dataset by an order of magnitude, and push GPT-5 into upstream feature extraction to consolidate stages. For applied teams, the lesson is that shipping a fully automated decision requires solving the explanation problem alongside the accuracy one, because trust, not raw precision, is what lets a human step out of the loop.