Skip to content

Newsroom

News and articles

Insights, updates, and engineering stories about AI, SaaS, security, and infrastructure.

Page 2 of 7

Latest posts

Analysis and engineering notes on AI, architecture, delivery, and search.

June 5, 2026

An Agent in CI Reopens the Boundary You Already Closed

The review reflex fires on checkout and stays silent on an agent reading a bug report

Agentic CI moved the attack surface from code execution to text: reaching a job that holds your secrets now costs one filed issue. Splitting the workflow by privilege closes that; a filter in front of the model only lowers the share of attempts that land.

Ihor K

CEO

Read article

June 2, 2026

Rate Limiting Is the Wrong Control for HTTP/2 Memory

One connection can pin more memory than a thousand well-behaved ones — budget per connection, at every hop

The HTTP/2 Bomb, disclosed on 2 June by Calif, chains HPACK indexed-reference amplification with a zero-window flow-control stall, so one compliant connection pins memory the server never frees. The control that answers it is a per-connection budget set at every hop that speaks the protocol; requests per second never touches it.

Ihor K

CEO

Read article

May 30, 2026

Fifteen Minutes of Cache, Then Three Hours of 404

Google handed the account back nine minutes after taking it, and Railway's edge went on answering 404 for services that had never stopped running

Railway's incident report puts an automated Google Cloud account suspension at 22:20 UTC on 19 May 2026, the first 404s fifteen minutes later, and edge traffic serving again at 01:38 — three hours in which services on AWS and bare metal were running and unreachable. The account itself had been returned at 22:29, before the first 404. What set the length of the blackout was what the edge proxies did with a cache they could no longer refresh.

Ihor K

CEO

Read article

May 12, 2026

Every Job That Can Write the Cache Can Publish

Eighty-four malicious @tanstack versions carried valid provenance: the signature named the right repository, the right workflow and the right commit, and every word of it was true

An attacker published 84 malicious versions across 42 @tanstack packages without stealing a publish token: a poisoned Actions cache, then an OIDC token read out of runner memory. Provenance verified correctly throughout, and that is the finding — the signature was accurate and the artifact was hostile.

Ihor K

CEO

Read article

May 1, 2026

When Egress Reprices, the Fix Goes Into Application Code

No setting chose this rate and none can unset it; the miss rate it multiplies is written in your handlers

Google raised list prices for data transfer out via CDN Interconnect, Direct Peering and Carrier Peering on 1 May 2026 — a rate that follows the destination of the packets, chosen by no setting in your stack and unsettable by any. Origin egress is the miss rate times bytes delivered times that rate, and the two factors you still control are both written in application code.

Ihor K

CEO

Read article

April 23, 2026

Decode and Training Now Ship on Separate Silicon

The serving part shipped with more memory and less arithmetic than the training part, and whether you can ever use it was settled by kernels written last year

Decode streams weights out of memory and does a few FLOPs per byte moved; training does the opposite, so a part built for both is tuned for neither. Google's 22 April announcement splits its eighth-generation TPUs along exactly that line, and the spec sheets confirm it: the inference part carries more HBM, more bandwidth and less peak arithmetic than the training part. Whether your stack can follow it is a property of your kernels, not of your contract.

Ihor K

CEO

Read article