Blog

Notes from the accountability layer.

The multi-model era is here. We write about what it changes: your costs, your security, and who answers for what your agents actually do. From the people building Gate.

RSS feed →
An indigo chain of linked circles crosses a dark navy field with one link missing, the severed ends marked in amber, and a dashed amber line rising from the gap to a checkpoint bar above.
N° 09 August 21, 2026

Delete the chat. The record survives.

A CEO deleted his ChatGPT logs. The court quoted them anyway. Keep your own record before someone else assembles one for you.

Orthogonal circuit traces on a dark blue ground. Every trace begins on a single amber vertical line at the left and routes rightward in right angles, branching as it goes. Most traces are indigo and a few are chartreuse, and each one ends in a small square pad.
N° 08 August 18, 2026

Software is writing itself

A $40 dev board that became whatever I described. A local model made four times faster by an agent with more patience than I have. And DeepSeek Harness, a program that rebuilds itself while you are using it. On software that stops arriving finished, what that does to a security team's idea of an approved application, and what has to hold still while everything above it changes.

Twelve rows of twenty small rounded squares on a dark blue ground. The top five rows are solidly filled with indigo squares, the rows below thin out from the left, and the bottom row holds three chartreuse squares among empty outlines.
N° 07 August 13, 2026

Frontier models fall to attacks you can script

OpenAI's whole GPT-5.6 line complied with an injected send in every trial, Grok 4.6 fell on the morning of its release, and Claude Sonnet 5 fell in the deepest session we ran. The payloads are short, fixed, and replayable, with no adaptive loop anywhere. On what a one-line send rule in the system prompt is worth against the 2026 frontier, which models hold, and what the same conversations do through Gate.

Three sets of curved strands, indigo, amber and grey, enter from the left at three different heights, cross and braid together, then leave to the right as eleven evenly spaced parallel indigo lines on a dark blue ground.
N° 06 August 7, 2026

AI Has Three Cost Centers. CFOs Need One Operating Model.

Every AI workflow now carries three costs at once: the tokens and compute it burns, the human work it reshapes, and the governance needed to explain the result later. One of those arrives as an invoice. The other two arrive as a surprise. On what a CFO should ask for, and why a verifiable record belongs inside the productivity model rather than beside it.

Seven rows of nineteen small rounded squares on a dark blue ground. The upper rows are almost solidly filled with indigo squares, lower rows thin out to scattered fills, and the bottom row holds only three chartreuse squares among empty outlines.
N° 05 August 6, 2026

Your system prompt is not a security control

We wired a defended email agent to a test inbox and sent it one forged message. Seven models followed it, five of them in every trial, and on a real self-hosted agent the model went looking for a mail tool it had not been given in order to comply. On what a written instruction to distrust tool output is actually worth, which models refuse this class outright, and what the same requests do through Gate.

Two ten-by-ten fields of small squares on a dark blue ground. The left field is filled to just under half with soft indigo squares; the right holds well under half that many, with the difference left behind as empty amber outlines, and a thin amber curve falls away from the taller field to the shorter one.
N° 04 July 28, 2026

The Redeployment Problem

AI is replacing tasks rather than whole jobs, and the companies furthest along are redeploying people instead of cutting them. But redeployment means taking the human out of the loop, and that requires being able to reconstruct what the AI did. On why audit infrastructure belongs in the productivity conversation and not just the compliance one.

Thin strands fan out from a single point on the left into steeply diverging grey and amber lines, while two indigo lines hold a lower, flatter path across a dark field.
N° 03 July 23, 2026

AI subscriptions will end

Most of us met AI through a flat monthly subscription, and that price was never the real cost of the compute underneath it. As work shifts from chat to agents, usage varies by orders of magnitude and serious AI spend moves to metered billing whatever the plan is called. On why the subscription era is ending, and how compression, caching, and model choice keep the bill from running away with the work.

Many request streams entering a single gate; one is stopped and blocked in red while the rest pass through and emerge as ordered, screened lines.
N° 02 July 14, 2026

Faster Isn't Efficient: Governing the AI Now Running Across Your Organization

Every team is adding AI, and almost none are adding visibility. Automation, efficiency, and transparency are three different things, and speed without a shared way to see, secure, and prove what your AI is doing isn't efficiency; it's the wrong outcome at scale. On the one layer that closes the gap across every AI client you run.

Many agent endpoints on the left and a few provider endpoints on the right, all routed through a single node labeled one gateway
N° 01 July 6, 2026

Pick your models. Own your rails.

Buy AI from a frontier lab and you get a whole operations department with the model: the screening you never see, the spend limits, the dashboard, the record. That bundle is coming apart as teams go multi-model. On why the accountability layer can't come from the people selling you the models, and what we built.