Blog
Insights & Updates
The latest on AI safety, product updates, and best practices from the Hyperfence team.
Oct 7, 2026
A decision you can argue with is still an argument
OpenAI's new Decisions API gets something important right about how AI should make choices. It also shows how far the idea still has to go.
By Keith Goatley Read more →
Hyperfence
Sep 30, 2026
How would you know if AI had slowed down?
Every major lab says AI should slow down. OpenAI has held back three times. Nothing on the public record can tell you whether any of it is pacing.
By Keith Goatley Read more →
Sep 23, 2026
Money doesn't buy a better answer. It buys a second opinion.
The largest studies of AI at work find it narrowing gaps, not widening them. But almost none of them asked whether the work was right. A subscription may buy a better first draft, but its clearest advantage is the means to check one.
By Oli Cundale Read more →
Sep 8, 2026
Whose safety policy is your AI following?
The Pentagon's AI shortlist was settled by which supplier would agree to the fewest limits. That is a procurement story, but it answers a question every organisation has - a safeguard you did not build, on infrastructure you do not run, under terms you did not write, was never yours.
By Simon Holliday Read more →
Aug 26, 2026
Memento: why your AI carries its own notes, and why anyone can rewrite them
A language model remembers nothing between turns. The whole conversation is sent back with every request, which means the conversation is something the client can edit, and a forged turn only has to read like the real thing.
By Simon Holliday Read more →
Aug 19, 2026
Why AI Models Refuse to Help Incident Responders
Frontier models refused to help Hugging Face's incident responders read the logs of an attack on their own infrastructure. That refusal isn't a bug and won't be patched - and it's about to hand a lot of security teams a governance problem.
By Simon Holliday Read more →
Jul 28, 2026
The Hyperfence demo is live: Don't talk about Fight Club
Our interactive Hyperfence demo is live. Watch requests get blocked before they reach the model, in milliseconds, try to make it talk about Fight Club.
By Tim Fraser Read more →
Jul 22, 2026
Where to Draw the Line - How We Build a Hyperfence
Anyone can write a keyword filter, and it'll be wrong constantly. Here's how we build a Hyperfence - a detector that catches a concept however it's phrased, while staying quiet on everything that only looks similar.
By Simon Holliday Read more →
Jun 16, 2026
Age Checks Answer Who. The Harder Question Is What
The UK is banning social media for under-16s and intimate AI chatbots for under-18s. The hard part isn't who's online - it's deciding what to block.
By Simon Holliday Read more →
May 20, 2026
Hyperfence beta is live: the safety layer AI needs
Hyperfence Beta is live. Why we built a deterministic AI safety layer, what it blocks, and how teams can start using it today.
By Tim Fraser Read more →
