AI-assisted RF Hacking with GNU Radio

The talk argues that software-defined radio is a good fit for generative AI. GNU Radio flow graphs are visual, open source, and built on Python, so a model can write them, run them, read the output, and iterate. The speaker says experiments that once took months of setup can now be tested in hours, with AI acting as a force multiplier that still needs a skilled operator at the controls.

The middle section covers a GNU Radio MCP server the speaker built. It lets an assistant detect attached hardware, ask which antenna port to use before transmitting, load, save and validate flow graphs, search blocks, and look up band allocations and common modulations offline. Templates let a team fix recurring generation mistakes once and share the result. The live demo was affected by network trouble, so the speaker showed an FM receiver build and described, rather than ran, a demodulator for a custom CC1101 protocol, a two-radio modulation link, and a known-key dewhitening attack against a Meshtastic-style payload. He also warned that transmitting on an unterminated port can destroy hardware.

The takeaway is that models are improving and some of these tool features are being built into them, but domain-specific tooling and checks still help. Practitioners should verify AI-generated RF code and keep a human validating results. The speaker pointed attendees to a flag-capture challenge running on the demo hardware.