Anthropic Glasswing has been everywhere this week.
After reading the coverage, the scepticism, and the counter-arguments, five things stand out to me....
Glasswing is Anthropic's new cybersecurity initiative.
They've used a frontier AI model to find previously-unknown vulnerabilities in browsers and operating systems we all use, and pulled together a coalition of rival tech companies to fix them.
Here are my five key takeouts.
- The hype is earned.
Whatever you make of the claims, this is one of the most polished product launches AI has seen this year.
Timing, narrative, authority. Worth paying attention to how they did it.
- Anthropic appear to be doing this responsibly.
They found something that could be genuinely dangerous in the wrong hands.
Instead of just shipping, they shared results with affected companies first and issued credits to help fix it.
That's what responsible AI looks like when it's done well.
- We need proper AI governance.
A frontier model just found thousands of vulnerabilities that humans missed.
That's useful today, in the right hands.
The question is what happens when it's in the wrong ones.
The pace of this technology is outrunning the frameworks we have around it.
- The scaffolding matters more than the model.
A smaller lab called Aisle replicated Glasswing's results with a cheaper model the same week.
The value is in the prompts, the agent loop, the tooling, the domain data you wrap around it.
That changes what you should own vs what you should bring in.
- Think about your own security.
Don't wait for someone else to pen test your systems with AI.
Get hold of a model. Use existing tools to probe your own defences.
The cost of testing is dropping faster than the cost of getting breached.
7DOTS can help if you need us.
What's your take?
P.S. Anthropic named Project Glasswing after the glasswing butterfly (Greta oto), a creature known for its nearly transparent wings.
First published on LinkedIn.
