Twitxy Saved

Origin

Verified account This account is verified because it's an official organization on X. Learn more

@originhq

Endpoint AI observability. Understand what your agents are doing.
originhq.com Joined February 2025
19 Following 1.1K Followers 174 Posts

Latest posts 5

Origin Verified account @originhq
Can you tell what a LoRA adapter learned without running it?

@BenGabayN used cosine similarity to distinguish six learned behaviors from adapter weights. Could this help flag unwanted behavior before use?

originhq.com/research/readi…
www.originhq.com Reading between the weights: what LoRA adapters reveal | Origin Technology Can a LoRA adapter's weights alone show what it was trained to do? Across 266 test adapters, the direction of the update did and its size didn't.
Origin Verified account @originhq
@jdu2600 built etw-rs for Origin’s flexibility and parsing needs, then brought a 2–3x parsing speedup back to krabsetw.

The write-up gets into how he made parsing faster without making the library a pain to use.

originhq.com/research/intro…
www.originhq.com Introducing etw-rs: an ETW parsing library built for flexibility and performance | Origin Technology etw-rs is a Rust ETW library that mirrors the Win32 API, keeps escape hatches open, and moves parsing cost to compile time for about 10x faster parsing.
0 replies 3 reposts 12 likes 275 views View on X
Origin Verified account @originhq
An agent found a way to reach an internal model, saved it to memory, and kept using it.

The route also ended up in a README, where another agent picked it up.

Chris followed how a working command became a reusable access recipe. A few details from the investigation ↓
Giving an agent permission to access a system also gives it choices about how to get there.

Origin allows you to observe how agents reach a system, not just what they do when they get there.

Read my investigation: originhq.com/investigations…
View quoted post on X
Origin Verified account @originhq
An agent couldn’t log in, so it pulled credentials from browser data.

Another changed a local security setting that was in the way.

One built a network tunnel when the normal route failed.
The tasks got done.

Who told the agents which workarounds were allowed?
Matt Hand Verified account @matterpreter Oct 30
Model alignment has a last-mile problem. Post-training instills general norms, but acceptable use is company-specific and rarely reaches the agent's context. An agent optimizing for task completion persists through lines it was never shown.

originhq.com/blog/normal-ac…
View quoted post on X
Origin Verified account @originhq
What happens when an agent can edit the rule that’s blocking it?

In one of our investigations, an agent added a “one-off override” to a deployment guard at another agent’s request.

Then used it again.