I've been building two new sites over the last few weeks, and they go together, so I'm sending them to you together.

AI Hacking Incidents

The first is hackingincidents.ai. It tracks cyberattacks since 2023 where attackers used AI, or are credibly suspected of using it, based on published reports. Each incident is graded by evidence (confirmed, strong, or suspected), along with which model was used, what the AI did, and the sources.

The hackingincidents.ai analysis page

hackingincidents.ai

There are 286 incidents in it right now. In 2024, none of the 27 reported attacks describe AI running the attack. In 2026 so far, 78 of 174 do.

Offsec Harness Index

The second is offsecharnesses.ai. It tracks the open-source offensive AI agent harnesses we can verify on GitHub, and shows which ones attackers have actually run in real operations. Every attack row links to the vendor, IR firm, CERT, or government report it comes from.

The Offsec Harness Index

offsecharnesses.ai

It covers 154 harnesses so far. 14 of them have shown up in real attacks, used by 26 different attacker groups, and 7 are tied to confirmed breaches.

I built these because I wanted to know how much attackers are really using AI, and which tools they use to do it. The harness index links straight into the incidents, so you can go from a tool to the attacks it showed up in.

Both get updated as new reports come out. We've definitely missed some and gotten some wrong, so if you see something off, corrections are welcome on both.

Talk soon!

Daniel

In partnership with

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.