A backdoored Anthropic extension, the end of the coding interview, and junk-grade AI debt

The Pulse — reports that Anthropic has removed the DSA-style coding interview from its engineering hiring loop, incident.io has cut AI spend from $4,500 per developer per month in August to $1,600, Amazon has written AI usage into engineer and manager expectations after finding it correlated with productivity, and Microsoft has offered 60,000 employees the chance to leave by year-end. Hiring criteria, spend levels and performance expectations are all being rewritten at once by the biggest names in the industry — these are the answers your board will expect you to have when it asks where you stand. The Pragmatic Engineer

Atlassian — has published the numbers from its own agent-driven development rollout: issue cycle time down from 11.5 days to 6.7, agent loops in Jira automatically resolving 51% of potential vulnerabilities, and agents with access to its Teamwork Graph context engine returning 44% more accurate results on 48% fewer tokens. These are first-party figures to benchmark your own transformation against — and the playbook's core claim is that context plumbing and measurement, not coding speed, are what actually gate the results. Atlassian

Proton — has cancelled company-wide 're-interviews' in which staff would have had to prove how much they use AI, with dismissal implied for falling short — days after 404 Media reported on CEO Andy Yen's AI push, a culture of fear, and his dismissal of an employee's request to pentest new AI-written code before shipping it. A live case of the mandate arriving before the measurement: staff revolted, users cancelled subscriptions, and the same criteria are said to be moving into performance reviews instead. Decide how you measure AI use before you require it. 404 Media

Imprint Engineering's verification loop — requires every pull request to carry a verification file — what the agent harness actually checked, and what it could not — before the PR can leave draft, and each major code area wakes up with 4-6 agent-raised pull requests to review. A separate feedback tool catches failures above the level of a single PR, and the gaps the files expose become the roadmap for making agents self-drive. If you are scaling agent-written code faster than reviewer capacity, this is the most concrete, copyable process description published so far, including the honest tension between centralising it and letting teams explore. The loop runs overnight, improving the codebase while nobody is at a desk. Irrational Exuberance (Will Larson)

Claude vulnerability research — Anthropic says an internal research model built working exploit chains for Chrome, Safari, Android, Windows and the Linux kernel with little human help; Google assigned 62 CVEs from the Chrome findings alone, including 13 sandbox escapes. One Linux flaw, CVE-2026-98432, gave remote root, and Go programs have enabled the vulnerable MPTCP path by default since Go 1.24. Epoch AI separately got an open-weight model to compromise patched Firefox releases for about $1,000 each — specialist-team skills at API prices, so your patch latency now decides exposure. Anthropic

Claude Ultra — no longer applies the five-hour usage window to verified accounts on its $200-a-month plan, and Anthropic has renamed the tiers: the old Max 20x is now Ultra, the old Max 5x is just Max, with prices unchanged at $200 and $100. For seat budgeting this makes the top tier predictable — heavy users no longer stall mid-session — but only once accounts are verified, so build the verification step into any rollout. explainx.ai

Azure Deployment Environments — shuts down on February 22, 2027 with no successor product, and Microsoft Dev Box follows on September 18, 2028 — Microsoft's own FAQ says the recommended path, Windows 365, is not one-to-one, and remaining Dev Box workloads will be deleted after retirement. The capability going away is governed self-service: Bicep and pipelines rebuild the provisioning but not the catalog, portal and role rules, so platform teams that adopted it have about sixteen months to rebuild that layer or move to Backstage-style tooling. InfoQ

Anthropic — is expected to receive a double-B credit rating — the top tier of junk grade — after its IPO, according to Blackstone, the anchor investor in the $63 billion deal financing its chips, reversing its earlier expectation of investment grade before 2027. Anthropic and Blackstone both declined to comment. Lenders are pricing risk into your main agent vendor's spending model; worth reviewing contract terms and exit clauses at your next renewal. TechCrunch

Intel — shares lost almost 29% of their value on Friday, the steepest one-day fall in the company's modern history, after management guided investors to expect less than they had forecast. The slide erased roughly $12 billion from the value of the stake the US government holds in the chipmaker. Semiconductor volatility is now part of the AI cost base your board keeps citing; treat hardware and capacity price assumptions as moving, not fixed. CNBC via Techmeme

Policy token — OpenAI, Google, Microsoft and Anthropic will honour a shared 'policy token' inside robots.txt that lets a site declare its AI scraping preferences once for all of them instead of crawler by crawler. Anthropic already reads it, Google and Microsoft plan support next year, OpenAI within months. One file now sets training and grounding policy across most major vendors — if you publish content or public documentation, add it to your web standards. The Verge

OpenAI — has cut its funding target to about $20 billion at a $1.25 trillion valuation after investors trimmed commitments — a first crack in bottomless AI capital. Financial Times via Techmeme

Big tech capex — Alphabet, Meta, Microsoft, Amazon and Oracle have committed over $3 trillion to AI infrastructure — the spending wave your cloud bills will repay. Bloomberg via Techmeme

StackOverflow — now gets fewer than 1,000 questions a month — the shared knowledge base your agents trained on has stopped regenerating. data.stackexchange.com · from the Hacker News discussion