Ephemeris
“935 attacks crossed a terabit per second. Most of them landed in the second quarter.”
Story 01 · Cloudforce One, DDoS Threat Report H1 2026
935 attacks crossed a terabit per second.
Cloudforce One's half-year report puts numbers on something most capacity plans still treat as an outlier. 23.2 million network-layer attacks mitigated in six months — about 5,343 an hour — plus 29.64 trillion HTTP DDoS requests blocked. The vector mix moved too: DNS floods went from a quarter of network-layer attacks to two fifths in a single quarter, and CLDAP reflection climbed 580% to take third place.
$ cloudforce-one report --period h1-2026 network-layer attacks ..... 23,200,000 (~5,343/hr) http ddos requests ........ 29.64T blocked attacks >= 1 Tbps ......... 935 (+519% q1 → q2) vector share, network layer, q2 dns flood ████████████████ 40.0% (was 25.7%) cldap reflect ████ +580% qoq, #3 vector dns, h1 total █████████████ 34.3% most-hit sector media, production & publishing 14.2% notable mover government #29 → #9 (operation epic fury)
Copilot will now call a model that never leaves your laptop.
Ollama joins Copilot for JetBrains as a bring-your-own-key provider, so the model picker in your IDE can point at something running on localhost. The same release adds Copilot Memory — a toggle in the settings portal that carries project facts between agent chat sessions instead of making you restate them every time.
11 Aug 2026
Copilot · BYOK
RE: what shipped
- Ollama as a BYOK provider — provider config and model selection throughout the JetBrains experience.
- Copilot Memory — retains and recalls context across agent chat sessions; toggled in the settings portal.
- Server-side admin controls for plugin availability and MCP server access.
- Copilot CLI auto-install from the integrated terminal on macOS, Linux and Windows.
One command, and your docs answer questions.
AI Search stopped requiring you to assemble the primitives yourself. Point it at a website or a folder, and you get crawling, chunking, embedding and hybrid retrieval behind a single instance — with a public /search and /mcp endpoint your agent can call without an auth dance.
- 1Sitemap-optional crawling — the Discover parsing option walks a site that never published one.
- 2Hybrid search: semantic and keyword matching in the same query, across multiple instances.
- 3Public endpoints on a custom domain, optionally fenced behind Cloudflare Access.
- 4Embedding and reranking are free on the default models. Ingestion $0.75/1M tokens, storage $2.00/GB-month, semantic queries $0.75/1k — roughly $35 a month in their worked example.
npx wrangler ai-search create my-search \ --source https://my-website.com \ --type web-crawler --hybrid-search
65.1% of the spend, on 30% of the tokens
Vercel's AI Gateway index for July is the closest thing to a public price sheet for how teams actually buy inference. Volume grew 59% and spend grew 37%, which means the average price paid per token fell 13.6%. Underneath that, the mix changed: DeepSeek took roughly a quarter of all gateway tokens and passed Google, with V4 Flash alone running more tokens than every Google model combined. Anthropic kept the money — 4.4× the average price, and two thirds of the spend.
4.9 trillion tokens, with the repositories left intact.
The Stack v3 is the largest open corpus of source code published so far, and the interesting change is structural rather than numeric. v2 shipped file identifiers against the Software Heritage graph; v3 crawls GitHub directly and stores file contents inline, so a training run sees whole repositories with their cross-file relationships still attached — which is exactly what an agentic coding model has to reason over.
Specification
- Train split
- 15.9 TB · ~4.9T tokens
- Coverage
- 713 languages · 173M repos
- Full split
- 113.7 TB · 770 languages · 224M repos
- Snapshot
- 7 August 2025
- Licence
- ODC-By v1.0
What changed since v2
Scale first: v2's training split was 2.0 TB and roughly 550 billion tokens. v3 is 15.9 TB and about 4.9 trillion — close to an order of magnitude, with two more years of open source behind it.
The pipeline is new as well. Licence detection runs through ScanCode, deduplication is language-agnostic via MinHash signatures rather than per-language, and StarPII scrubs personal information before release.
Attribution is the cost of entry: ODC-By permits commercial and non-commercial use provided the source is credited.
That's today.
Five picks, and three of them are prices: what a terabit-scale attack now costs to absorb, what a token costs in July, and what a search index costs a month. The other two hand you something to install and something to train on. Issue 111 · Wednesday 12 August 2026 · Zürich.
Filed from — Cloudflare · GitHub · Vercel · Hugging Face, surfaced via The Batch. Also read: The Zvi · Import AI · Jeff Bullas · OpenAI · Anthropic · Sentry · PostHog · Fly.io · Next.js · @seeallochnaya · @TochkiNadAI · @denissexy · @rvnikita_blog · @ProductsAndStartups.
Rubric — tools you could adopt this week, creative software, dev tools and agentic coding, privacy and security when the signal is high, research with a practical kernel. Nothing older than thirty days, nothing shipped here before.