Багхантинг через LLM: как специалист «Бастиона» ищет уязвимости и не тонет в спаме

Why an AI agent drowns in spam

Running an AI agent on a bug bounty program takes five minutes, but turning its "stream of consciousness" into accepted reports is the tricky part — one that can easily get you banned for spam and cost you your rating. The author has been bug hunting for about a year and a half, won the BugZone 7.0 competition, and for the past six months has been hunting bugs specifically through LLMs — in his words, AI can help you achieve results instead of submitting garbage.

*Author — specialist in the incident response department at Bastion, ~1.5 years of bug bounty experience

*Winner of the BugZone 7.0 competition

*Has been searching for vulnerabilities with LLMs for the past ~6 months

*The breakdown covers 4 approaches to AI bug hunting and 4 ways to pay for the model

*Three-layer verification and a pipeline of 2 AI agents, live walkthrough of a blind SSRF

How not to drown in false positives

The key to working AI bug hunting is multi-layered verification: the author builds three layers of verification that every model finding must pass before submission. Without this filter, bug bounty platforms treat raw LLM output as spam, which risks a block and a penalty to the researcher's rating.

Roles in the chain are split: one agent generates vulnerability hypotheses, the second verifies and sorts them, taking the routine of initial triage off the human's plate. It's precisely the manual review of the flow of false positives that, as the author notes, turns a researcher into a "free triager for their own agent" — which is exactly what needs to be avoided.

Separately, the author breaks down the economics: he lists four ways to pay for the model, because with mass hypothesis runs, token costs become a significant line item in a researcher's expenses.

What the blind SSRF breakdown shows

As a live example, the author walks through the search for a blind SSRF (server-side request forgery) vulnerability class — when a server, on an attacker's command, reaches out to internal resources but does not return the response directly. This case shows how the AI agent formulates a hypothesis, and the verification layers confirm exploitability before the report is filed.

«Running an AI agent on a bug bounty program is a five-minute job.

But turning its stream of consciousness into accepted reports, so you don't get banned for spam and lose rating points, is the tricky part.»

—

Vladislav, specialist in the incident response department, Bastion

The author doesn't promise a ready-made "make it good" button: even with an LLM, the result rests on the discipline of verification, not on automation for automation's sake.

What this means

LLMs already give ethical hackers a real edge in bug bounty, but only in tandem with a human and rigorous verification. A bare AI agent without verification layers generates spam that costs rating and program access — the value comes from the pipeline's architecture, not the model itself.

Frequently asked questions

What is AI bug hunting?

It's the search for vulnerabilities within bug bounty programs using large language models (LLMs): the model generates vulnerability hypotheses, and the researcher verifies and formats the findings. The specialist from Bastion describes 4 approaches to this process.

How does the AI agent keep the researcher from becoming a spammer?

Through verification: the author runs every model finding through three layers of checks and a pipeline of two agents, where the second agent takes on sorting the findings. Without this filter, raw LLM output leads to a spam ban and a rating penalty.