Building an Agentic Bug Bounty Hunter on a Raspberry Pi 5

10 hours ago

Mass automation in bug bounties can lead to noise without proper target understanding and tuning.
Agents are used as quality gates to overcome automation brittleness and improve recon data quality.
A tiered model approach is employed: Opus for strategy, Sonnet for execution, Haiku for lightweight tasks, and deterministic workers for non-model tasks.
The orchestration loop involves Python-controlled decisions, with Opus orchestrating actions like recon, test, authenticate, research, note, or done.
A knowledge graph system is implemented to build relationships between findings, endpoints, and tech stacks for better decision-making.
Semantic similarity via pgvector helps in reusing past findings and knowledge, enhancing the system's learning over time.
Custom tooling and strict role-scoped tools ensure controlled and efficient operations.
Epochs and timeouts are used to manage runs, allowing for comparison and improvement tracking.
Bounded context snapshots and queueing mechanisms ensure focus and efficiency in operations.
The system includes hardware like a Pi 5 with NVMe SSDs and an e-ink display for real-time status monitoring.

Hasty Briefsbeta