This AI pet doesn't play games; it captures WPA handshakes

What it does

Pwnagotchi is an A2C-based agent that uses an LSTM with MLP feature extractor to drive bettercap and capture WPA key material. It passively sniffs or actively performs authentication and association attacks, dumping handshakes—including PMKIDs, full and half WPA handshakes—into PCAP files for hashcat. Over successive epochs, it adjusts its own parameters to improve its yield in the specific RF environments you expose it to.

The interesting bit

The project deliberately rejects the usual gym-game sandbox for reinforcement learning and applies it to real-world RF chaos instead. Multiple units can even gossip over a custom parasite protocol built on top of 802.11 information elements, eventually learning to divide channels cooperatively rather than stepping on each other.

Key highlights

• Uses an A2C agent with an LSTM policy network to tune WiFi attack parameters over time.
• Captures PMKIDs, full and half WPA handshakes into standard PCAPs for offline cracking.
• Epoch duration scales dynamically from seconds to minutes based on local AP and client density.
• Nearby units detect one another via a custom dot11 parasite protocol and learn to share spectrum.
• Behaves like a pet: it gets “bored” in familiar environments and improves when exposed to novel networks.

Caveats

• Early epochs are explicitly an exploration phase; initial performance is intentionally poor while the agent figures out the local RF landscape.
• The unit requires physical exposure to new WiFi environments to keep learning; a stationary device will stagnate.

Verdict

Anyone looking for a hands-off WiFi auditing companion that improves with travel will find the novelty appealing. If you need deterministic, immediately optimal WiFi auditing, stick to static tooling.