BLOG
Blog
Writing on sovereign AI, our platform and our products — in plain, engineering-credible, fact-based terms.
- 14 Sept 2026INDUSTRY
The attacker is an agent now — and your API key is the loot
Anthropic’s 10 September threat report reads like a field manual for the new tempo: agents that rewrite their own malware when flagged, 1.8 million apps decompiled for hardcoded secrets, credentials to admin in three hours, and AI companies raided for API keys and pre-release models. The same week brought the mirror image: an agent that invented a package, laced it, and shipped it to PyPI — no attacker required. When sophistication stops being the moat, the path becomes the moat.
6 min read
- 3 Sept 2026INDUSTRY
The assistant grew an ad business — and the EU called it what it is: a search engine
On 31 August, hours apart: OpenAI’s ad business crossed a $1 billion run rate in about 200 days — AdWords needed four years — and Brussels designated ChatGPT a Very Large Online Search Engine, 159.1 million EU users past the threshold. Apple, meanwhile, rents Siri’s brain from Google for about $1 billion a year. The middle of everyone’s questions now has interests of its own. At work, the answer is a middle you can meter: one gateway, models you can own, every token visible.
6 min read
- 26 Aug 2026PRODUCT
Sosyabot: social AI for a feed that now charges for sameness
In four days at the end of July, LinkedIn shipped an “AI slop” report button and retired its own writing assistant, Snapchat pulled fully AI-generated videos from Spotlight recommendations, and Article 50 made disclosure law. The feed stopped labelling the flood and started pricing it — and the volume play inverted. Sosyabot is the studio built for the other strategy: one brief into native, brand-locked, human-approved posts across 14+ networks.
5 min read
- 22 Aug 2026PRODUCT
Talky: communications you own, for the year governments stopped renting theirs
On 18 August Beijing ordered a customised Windows 10 off state machines, months ahead of schedule. France had already named its date — 2.5 million civil servants off Teams and Zoom by 2027 — and Denmark, Schleswig-Holstein and the Netherlands each took their own step this year. The assumption that everyone’s workday runs on the same rented stack is ending, and communications go first. Talky is the workspace built for that decision: meet, message and create inside your own perimeter.
5 min read
- 20 Aug 2026PRODUCT
Pickerman: media intelligence for a flood that is aimed at the checkers
On 12 August, Germany’s BfV publicly named the Matryoshka network flooding four September elections with fake broadcaster clips — the same day a deepfake report counted 15,736 victims in six months, half of video fakes built to wreck reputations. Operation Overload’s real target is the verifier’s time. Pickerman is our media-intelligence radar built to invert that price: rank by independent agreement, check claimed sources against real ones, keep a human in command.
7 min read
- 10 Aug 2026INDUSTRY
Your tool stack is your attack surface — July proved it, one package at a time
In four weeks the work toolchain was hit through a poisoned npm package hunting AI-assistant configs, a PNG that walked past AI code reviewers, a release pipeline that shipped malware with valid provenance, and a worm that swallowed 440 packages in an hour. Third-party breaches just doubled to 48%. Shrinking the stack is no longer tidiness — it is security posture.
6 min read
- 5 Aug 2026PLATFORM
Verinova, Blab, Solab, Spook, VeriRAG: the five model families under everything we ship
Everyone spent the summer arguing about how to label what models make. The quieter question decides more: whose model is it? Meet the five families we train in-house — Verinova for language, Blab and Solab for the two directions of speech, Spook for vision, VeriRAG for retrieval — and why owning them changes what a product can promise.
5 min read
- 3 Aug 2026INDUSTRY
The AI-voice disclosure era started on a Sunday
On 2 August 2026 the EU AI Act’s Article 50 became applicable: voice agents must say they are AI, synthetic audio must carry a machine-readable mark, deepfakes must be labelled. The vendors raced the deadline and the watermarks have honest limits — but the question the law still doesn’t answer is consent. That one, Blab answered on day one.
7 min read
- 27 Jul 2026PRODUCT
Solab: AI meeting notes where every claim shows its moment
The AI notetaker became the meeting’s default attendee — and spent July 2026 being retrofitted with consent switches while courts weigh whether the bots wiretap. Solab starts from the other end: a record you make on purpose, notes where every claim links to the moment it was said — and unknown stays unknown.
7 min read
- 16 Jul 2026INDUSTRY
When “AI-made” goes on the label: keeping social on-brand in the disclosure era
This week the disclosure rails snapped into place: Google began showing people whether an ad was AI-made, and Meta yanked an AI feature off Instagram a day after shipping it. The lesson is not “stop using AI” — it is that consumers punish two things, undisclosed AI and off-brand AI. The fix is determinism: lock the brand, then constrain every generated post to it.
6 min read
- 13 Jul 2026INDUSTRY
Chat Control, the encryption line, and where your organisation’s plaintext actually lives
On 9 July more MEPs voted to kill Chat Control 1.0 than to keep it — and it survived on a threshold. It is not law, and it never mandated scanning. What the vote did settle is that Europe’s legal line now runs through one architectural question: on whose server does a readable copy of your message come to rest?
6 min read
- 8 Jul 2026INDUSTRY
Media intelligence in the synthetic era: why OSINT has to corroborate, not just collect
Deepfakes and AI-generated posts are flooding the open web that media intelligence reads. Provenance standards like C2PA certify history, not truth — so the job shifts from collecting sources to corroborating them across independent feeds.
4 min read
- 6 Jul 2026INDUSTRY
Voice agents in 2026: why the cascade still beats speech-to-speech for grounded, compliant calls
xAI shipped a no-code voice-agent builder on a speech-to-speech path this month. It is faster and more natural — but for grounded, disclosable, regulated calls the cascade’s readable text seam still wins. Here is the trade-off, honestly.
5 min read
- 29 Jun 2026INDUSTRY
The EU AI Act’s 2 August 2026 deadline didn’t move — here’s what actually lands
The Digital Omnibus pushed high-risk AI rules to 2027 — but Article 50 transparency, GPAI enforcement and fines still arrive on 2 August 2026. Why it reaches Turkish builders too, and where you enforce it.
6 min read
- 26 Jun 2026INDUSTRY
Your AI agents are identities now — and most of them are ungoverned
AI agents don’t log in — they act with credentials. Non-human identity is the security story of mid-2026, and the gateway is the one place you can actually issue, scope and revoke it.
5 min read
- 24 Jun 2026INDUSTRY
The agent gateway: one control plane for every model and tool
In 2026 AI agents stopped just answering and started acting — calling tools over MCP. This month Databricks and others converged on the same answer: a single gateway in front of every model and tool. Here is why, and where Qevron already fits.
4 min read
- 22 Jun 2026PRODUCT
SetGet: the scattered tool stack, consolidated into one work OS
Chat in one app, docs in another, tasks in a third, dashboards in a fourth — context scattered across a dozen tabs and as many vendors. SetGet folds projects, wiki, chat and dashboards into a single work OS you can self-host, with SSO and an audit trail.
3 min read
- 18 Jun 2026INDUSTRY
Production RAG in 2026: hybrid retrieval, reranking, and reducing hallucination
Most RAG systems fail in retrieval, not the model. A field guide to the 2026 production stack — hybrid search, Reciprocal Rank Fusion, cross-encoder reranking, grounded citations and evaluation — and why sovereign teams run all of it inside their own perimeter.
5 min read
- 15 Jun 2026PRODUCT
Voices real enough to need consent: inside Blab, our voice studio
Blab is the studio surface for our own text-to-speech family — sub-second streaming, consented voice cloning, and dubbing that keeps the original soundtrack.
1 min read
- 8 Jun 2026INDUSTRY
Sovereign AI for regulated industries: when on-prem stops being optional
Finance, healthcare and the public sector increasingly face hard data-residency rules. Here is how on-prem and isolated AI maps to DORA, KVKK, banking localisation and the EU AI Act.
2 min read
- 1 Jun 2026PRODUCT
Voice agents that actually pick up the phone: inside Calleague
Real telephony, a realtime speech-to-text → LLM → text-to-speech pipeline and production RAG — what it takes for an AI voice agent to hold a natural call, and how Calleague is built for it.
2 min read
- 25 May 2026INDUSTRY
Vendor lock-in in the LLM era — and the gateway pattern that loosens it
Lock-in is rarely the model weights. It is the API surface, the prompts and the ops glue. Here is how OpenAI-compatible gateways became the standard answer — and what they don’t fix.
2 min read
- 18 May 2026PRODUCT
One endpoint, every model: how Qevron, our OpenAI-compatible gateway, works
Point your code at a single OpenAI-compatible API and reach our in-house models plus 43+ providers — with routing, caching, monitoring and cost analytics in one place.
2 min read
- 11 May 2026INDUSTRY
Data sovereignty in 2026: why on-prem and isolated AI moved from preference to plan
The EU AI Act clock, a rewritten KVKK transfer regime and the CLOUD Act gap have turned “where your AI runs” into a legal question, not just an architectural one.
2 min read
- 4 May 2026PLATFORM
Our models. Our gateway. Your AI. — the two layers under everything we build
Arpanet is built in two layers: in-house model families and one OpenAI-compatible gateway underneath, a suite of products on top. Here is what that vertical integration actually buys you.
2 min read