// news & analysis
GPT-6 Astra Attacked Out-of-Scope Targets in 29% of UK AI Security Institute Tests
The UK AI Security Institute found on September 28 that GPT-6 Astra completed unsanctioned supply-chain attacks in 29.2 percent of…
Category
Category: Production AI
Hands-on engineering for AI systems that run in production: architecture, reliability, data, and security.
// news & analysis
// news & analysis
Claude Opus 5.5’s 40% Price Cut Is Half List Price and Half Lower Default Reasoning
Grok 4.7, Claude Opus 5.5 and GPT-6 Sol and Luna all landed on September 21 and 22, with cuts of…
// news & analysis
Cloudflare Containers Leaked Residual Disk Data Between Tenants on 18 of 24 Placements
Cloudflare disclosed on September 26 that residual disk blocks in Containers and Sandboxes could be read across tenants, on 18…
// news & analysis
Microsoft’s Own Data Shows Copilot Cut Click-Throughs to the New York Times by 93%
Unredacted summary judgment briefs in the Times case put a measured number on AI substitution: Microsoft's own telemetry showed Copilot…
// news & analysis
Anthropic Is Paying Akamai $11.6 Billion for the CPU Compute Around Its Models
Anthropic is paying Akamai $11.6 billion over seven years, expandable toward $20 billion, and the compute is CPU rather than…
// news & analysis
Cognition’s $47 Billion Valuation Depends on Models From Its Own Competitors
Cognition is raising about $1B at a $47B valuation and everyone called it the frothiest number of the year. Run…
// news & analysis
Google and OpenAI Put Their Strongest Security Models Behind Private Allowlists
Google gated Gemini 3.8 Flash Cyber behind a 650-organization allowlist on September 2. A day later OpenAI shipped GPT-6 Astra,…
// news & analysis
OpenAI Is Cutting Off Cursor’s Model Access Because of a Change in Ownership
OpenAI is terminating Cursor's model access on November 12 — not over price, not over policy, but over ownership. Four…
// production ai
How to Evaluate AI Vendors as a Technical Buyer
Verifying the architecture is necessary but not sufficient. The things that sink an AI vendor relationship in year two are…
// production ai
How to Verify Which Architecture Sits Behind an “AI-Powered” Product
Underneath every "AI-powered" label sits one of four architectures, each with a different cost curve, latency profile, accuracy ceiling, and…
// production ai
What I Learned Rebuilding the Ingestion Layer of a Production Document Agent
Document ingestion is the number one pain point in small-business operations, and most teams blame the model or the data…