Troubleshooting AI Bug Hunting: How Not to Become a Free Triage for Your Agent

Launching an AI agent for bug bounty takes minutes, but without verification it produces a mass of false positives, turning the researcher into a free triager. For example, out of 67 reports, only 3 may be valid. According to HackerOne data, the volume of incoming reports in 2026 grew by 76% year over year, while the share of confirmed exploitable findings remains at 25%, but the share of critical and high findings increased to 32%. At the same time, the growth of valid AI reports was 210%. Platforms are introducing restrictions: curl closed its program on HackerOne as of February 1, 2026; HackerOne suspended submissions for the Internet Bug Bounty; BI.ZONE and Standoff 365 introduced daily report limits; and Brave bans users after two spam reports. A study (Zheng et al., 2025) on 9,942 reports found that XSS constitutes 13.8% of submissions, Information Disclosure 9.4%, and Access Control with Authentication over 10%. Modern LLMs tend to accept invalid reports, but RAG helps solve this problem. According to Bugcrowd surveys, 82% of researchers use generative AI in 2026. The author identifies four approaches to AI bug hunting: manual analysis with an LLM advisor, auto-scanners with LLM interpretation, a boxed agent, and a verified pipeline. Models are paid for using local setups (Ollama, LM Studio, vLLM), APIs (powerful models), subscriptions (Cursor, Claude.ai, ChatGPT), or enterprise platforms (Bedrock, Azure OpenAI, Vertex AI). The architecture for verifying findings includes three layers: agent rules (e.g., time-based check for SQL injection, out-of-band for server-side request forgery, headless browser for cross-site scripting), RAG (access to a knowledge base with vulnerability patterns), and model context protocol (tools for active hypothesis testing).