PostgreSQL 19 and PHP 8.6 both landed on Thursday, Valkey 9.2 took the fork() penalty out of database snapshots, and Microsoft used Ignite to make Azure HorizonDB and its Cobalt 200 Arm virtual machines generally available. Chrome 158 also switched off XSLT in the browser after years of XML parser bugs.
On the model and market side, OpenAI shipped GPT-6.2 Astra and Sol, Google slotted Gemini 4 Neon between Hydrogen and Carbon, and Anthropic debuted on Nasdaq as ANT in a $93.5 billion public offering paired with a $26 billion chipmaker placement. Security teams spent the same week patching Java4Shell in OpenJDK, HTTP/2 flow-control bugs in NGINX, and two poisoned releases of Microsoft’s Python agent package on PyPI.
Top Stories This Week
PostgreSQL 19 Repacks Tables Online and Pins Query Plans Without Code Edits -
On November 19, the PostgreSQL Global Development Group released PostgreSQL 19. Busy tables that take constant updates and deletes accumulate dead space, and cleaning that up with VACUUM FULL or CLUSTER has always meant locking the table until the rewrite finishes. Most teams reached for third-party tools like pg_repack instead. PostgreSQL 19 builds the answer into core SQL: REPACK reclaims dead space and can reorder rows to match an index, and REPACK (CONCURRENTLY) does the job without blocking reads or writes.
Query plan flips get a built-in answer too. When table statistics drift, the planner can trade a fast plan for a slow one at the worst moment. The new pg_plan_advice extension lets you test and lock in plan choices using a compact plan-advice syntax, while pg_stash_advice attaches that advice to a query identifier so you can stabilize a bad production query without redeploying application code. Autovacuum can also clean indexes across parallel workers via autovacuum_max_parallel_workers, and ordinary SELECT queries can mark heap pages all-visible as they read them.
Several everyday queries get shorter or drop a network round trip. INSERT ... ON CONFLICT DO SELECT lets RETURNING hand back the existing row when an insert hits a unique constraint, covering the standard get-or-create pattern in one statement. Sessions reading from a replica can run WAIT FOR LSN to block until the standby has replayed a specific write-ahead log position, giving read-your-writes consistency without routing every follow-up read back to the primary. Turning on logical replication no longer needs a server restart either: creating the first logical slot raises effective_wal_level to logical on the fly, and CREATE PUBLICATION ... FOR ALL SEQUENCES replicates sequence values alongside tables. Foreign key checks are faster too, nearly doubling bulk insert speed in testing.
Check a few changed defaults before running pg_upgrade. Just-in-time compilation (JIT) is now disabled by default because the cost estimates that triggered it proved unreliable on short queries; turn it back on explicitly if you run heavy analytical scans. When built with LZ4, default TOAST compression switches to lz4. Server-side TLS Server Name Indication (SNI) is now supported through pg_hosts.conf, successful md5 logins emit warnings ahead of future removal in favor of SCRAM, and RADIUS authentication is gone.
PHP 8.6 Adds Partial Function Application, Native Stream Polling, and Safer Sessions -
On November 19, PHP 8.6.0 shipped on its regular late-November schedule. Passing a callback to array_map() or array_filter() has often meant wrapping a one-line function call inside an arrow function just to fix one or two arguments. Partial function application lets you call the function with ? and ... placeholders for the missing arguments and get back a Closure:
1
$modern = array_filter($versions, version_compare(?, '8.6', '>='));
Async PHP libraries get a lower-level upgrade with Io\Poll. Until now, pure-PHP event loops relied on stream_select(), which slows down and hits descriptor limits around 1,024 open sockets. Io\Poll\Context wraps epoll on Linux, kqueue on macOS and BSD, event ports, and WSAPoll behind one standard API that projects like ReactPHP and AMPHP can share. Stream contexts can also set StreamErrorMode::Exception so failed file and network calls throw a typed StreamException instead of emitting warnings that callers have to scrape out of error_get_last().
The standard library fills in several long-standing gaps. \Time\Duration measures elapsed stopwatch time independent of timezones, with nanosecond precision across more than 250 years. clamp($value, $min, $max) bounds numbers or any comparable object, including Duration. A built-in SortDirection enum (Ascending and Descending) stops query builders from accepting unrelated integer constants like SORT_STRING, readonly properties can now declare default values without boilerplate constructors, function parameters can carry their own DocComments, and the Sodium extension adds ML-KEM-768 and X-Wing post-quantum key encapsulation.
Upgrading from 8.5 brings tighter session defaults and a list of deprecations. session.use_strict_mode and session.cookie_httponly are now enabled out of the box, and session.cookie_samesite defaults to Lax. fclose(), file_put_contents(), and copy() return false when a buffered flush fails instead of hiding the write error. Using return inside a finally block is deprecated because it silently swallows active exceptions, and the mb_ereg* regex functions are deprecated because the underlying Oniguruma C library is unmaintained.
Valkey 9.2 Cuts Forking Out of Snapshots and Moves Sorted Sets to B+ Trees -
On November 18, Valkey released 9.2.0 GA. Anyone who has run a large Redis or Valkey primary under heavy writes knows the snapshot tax: background RDB saves and replica full syncs call fork(), and Linux copy-on-write page faults can double memory pressure while clients keep mutating keys. Valkey 9.2 adds forkless RDB snapshots, enabled with forkless-infrastructure-enabled and bgsave-default-method, plus forkless full synchronization that streams snapshots to replicas from a background thread without forking. RDB persistence and replication also add streaming Zstd compression alongside whole-stream rdbcompression lz4.
Under the hood, large sorted sets now sit on a B+ tree instead of a skiplist, shrinking memory overhead and speeding up ZPOPMIN, ZPOPMAX, and ZSCAN. Code that inspects OBJECT ENCODING will see btree where it used to see skiplist, and range commands like ZRANGEBYSCORE now reject a bare ( or empty string as a score bound instead of quietly treating it as 0. A new radix-tree data type called Path Hash (PH* commands) handles exact lookups, prefix walks, and longest-prefix matching on binary-safe paths.
Several sharp operational edges get first-class commands too. DELEX deletes a string key conditionally when its value matches (IFEQ) or differs from (IFNE) an expected token, replacing tiny Lua lock-release scripts. INCREX increments a counter and sets its TTL in one atomic call, XACKDEL and XDELEX clean up stream entries once consumer groups finish with them, ACL SETROLE defines reusable permission roles, maxmemory-scripts caps cached Lua memory, and setting hotkeys-top-k turns on built-in hot-key detection via HOTKEYS GET.
Microsoft Ignite: Azure HorizonDB and Cobalt 200 VMs Reach General Availability -
At Microsoft Ignite on November 17, Microsoft made Azure HorizonDB generally available across 13 regions. HorizonDB is Microsoft’s cloud-native PostgreSQL service with distributed storage up to 128 TB, multi-zone durability by default, sub-millisecond commits, and failover in under five seconds. It ships with pgvector, DiskANN vector indexing, pg_textsearch BM25 keyword ranking, and a tuned Apache AGE graph extension so teams can keep relational tables, hybrid search, and graph traversals inside one Postgres cluster. Alongside it, Microsoft Fabric HorizonDB entered public preview as a serverless Postgres tier that boots in seconds, pauses compute after five minutes of idle time, and mirrors tables straight into OneLake Delta format.
On the compute side, Azure Cobalt 200 virtual machines built on Microsoft’s second-generation Arm64 server chip reached general availability, with deployments expanding to 24 Azure regions by year-end. Microsoft says Cobalt 200 delivers up to 50% higher CPU performance than Cobalt 100, adds selectable consistent-speed and boost modes, and introduces memory-heavy Mpsv4 and NVMe-heavy Lpsv5 sizes with up to 23 TB of local storage.
For platform teams that want managed AI tooling without handing workloads to a multi-tenant cloud endpoint, Microsoft also previewed AKS AI Manager for Foundry. It deploys open-weight models and Foundry hosted agents onto AKS clusters you already run, isolating each agent session inside a lightweight VM pod sandbox while keeping the regular azd and Foundry API workflow. Azure SQL Database also added a free local development container in preview, Azure Cosmos DB made its native Rust SDK and VS Code shell generally available, and Windows Server 2028 opened in preview.
OpenAI Ships GPT-6.2 Astra and Sol With a 20% Price Cut on Astra -
On November 17, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in the API, ChatGPT Work, Codex, and Amazon Bedrock. Astra is the heavy model for multi-file coding, scientific terminal tasks, CAD and 3D web builds, and structured spreadsheets or slide decks, while Sol is the faster workhorse OpenAI points at everyday engineering and agent loops. OpenAI says Astra answers with 22% fewer output tokens on its internal knowledge-work suite than GPT-6 Astra, and Sol lands within half a point of Astra on Cognition’s FrontierCode 2.0 at roughly one-fifth the per-task cost.
The API model ids are gpt-6.2-astra and gpt-6.2-sol. Both offer a 1,050,000-token context window and up to 128,000 output tokens. Standard pricing for gpt-6.2-astra drops 20% below GPT-6 Astra to $8 per million input tokens, $0.40 for cached input, and $40 per million output tokens. gpt-6.2-sol keeps GPT-6.1 Sol’s standard rates of $2 input, $0.10 cached input, and $10 output (requests over 272K input tokens pay double on input and 1.5x on output). Both models work on /v1/chat/completions and /v1/responses, though tool calling requires the Responses API, and requests that set reasoning effort to none are rejected. For async jobs, a new Responses API public beta offers two-hour and eight-hour background completion windows at 30% and 40% off Standard token rates as Flex processing winds down.
OpenAI also marked GPT-6.2 Astra as its first externally released model to cross the High threshold for AI Self-improvement under its Preparedness Framework, alongside Critical for cybersecurity and High for biological capabilities. That classification comes with stricter automated classifiers that can block dual-use security testing outside OpenAI’s Daybreak program.
Google Slots Gemini 4 Neon Between Hydrogen and Carbon -
On November 18, Google introduced Gemini 4 Neon, filling the middle tier of the Gemini 4 lineup it started rolling out last week. Carbon handles the longest, costliest reasoning runs and Hydrogen handles cheap, high-speed turns; Neon is adapted from Carbon to run repository-wide coding, analytical work, and tool-heavy agent sessions at lower latency and about a 32% lower cost per task than Carbon at its default high thinking setting.
The API id is gemini-4-neon, available on the paid tier of the Gemini API in Google AI Studio, Android Studio, and Gemini for business, with no free API tier. Pricing is $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens across a unified 1,048,576-token window for text, images, audio, video, and PDFs. Synchronous calls cap output at 65,536 tokens, while setting background=True on the Interactions API unlocks long decode up to the full window, with thinking levels low, medium, high, and xhigh.
In Google Antigravity, Google also paired the new model with Hydrogen in an Orchestrator (Neon) mode for paid Pro and Ultra subscribers. Neon writes the plan and reviews diffs, while Hydrogen runs the edits, package installs, and test suites, handing control back to Neon only when a step gets stuck.
Developer Tools & Platforms
Copilot SDK 2.0 Moves to a Rust Runtime, and Vercel Ships eve 1.0 -
Running an AI agent inside your own product has usually meant starting a dedicated container for every user session, even when the user only asked a question against two HTTP APIs. On November 17, GitHub and Microsoft released Microsoft Copilot SDK 2.0 (formerly GitHub Copilot SDK) with a multi-tenant runtime rewritten in Rust that embeds directly in-process via a C ABI across TypeScript (@microsoft/copilot-sdk), Python, Go, .NET, Java, and Rust. On a warmed workload, spinning up a session and finishing one turn fell from 5.25 seconds on the old TypeScript CLI to 55.3 milliseconds, while memory for 10 clients dropped from 1,383 MB to 126 MB. Setting mode: "empty" strips host shell and disk access so sessions only see tools you allow, passes per-user tokens instead of a server credential, and writes session journals through sessionFs so any worker can resume a conversation or promote it into a container sandbox later.
The same day, Vercel made eve 1.0 generally available as an open-source framework for durable production agents. Scaffold a project on Node.js 24 or newer with npx eve@latest init my-agent, and add repo editing with eve add extension/code. Built on Workflow SDK 5.5, eve checkpoints only changed conversation deltas so runs survive process crashes without replaying completed steps, cutting mock-model median turn overhead from 3.04 seconds in September to 414 milliseconds. It ships with OpenTelemetry tracing, multiplayer Slack and web approval cards, and native route integrations for Next.js, SvelteKit, Nuxt, TanStack Start, and Vite.
AWS Open-Sources ExtendDB 1.0 and Doubles Lambda Timeouts and S3 Read Rates -
Teams that write against the DynamoDB API have never had an official way to run that same code on their own Postgres or SQLite database in production. On November 18, AWS released ExtendDB 1.0 under the Apache 2.0 license, a DynamoDB-compatible adapter backed by PostgreSQL, SQLite, Cassandra, or MongoDB. For local development and CI, the @extenddb/dev npm package and extenddb-dev container spin up a local DynamoDB endpoint over SQLite with no external database setup.
AWS followed over the next three days with a pre-re:Invent batch that moves several familiar platform ceilings. On November 19, AWS Lambda doubled the maximum timeout to 30 minutes for on-demand Event Functions (up from 15 minutes), added OpenTelemetry tracing for durable functions, and on November 20 added Node.js 26 and Python 3.15 runtimes plus 10,000-message batches for provisioned SQS FIFO mappings. The same day, Amazon S3 doubled its per-prefix read rate from 5,500 to at least 11,000 GET and HEAD requests per second, EC2 T9g burstable instances powered by Graviton5 became generally available, and CloudFormation raised its default stacks quota from 2,000 to 300,000 per account and Region.
Spring Cloud 2026.0.0 and Spring AI 2.1 Reach GA on Spring Boot 4.2 -
On November 20, Spring Cloud 2026.0.0 (Paddington) reached general availability on top of Spring Boot 4.2.0. Import the BOM org.springframework.cloud:spring-cloud-dependencies:2026.0.0 in Maven or Gradle. Spring Cloud Bus has been folded into Spring Cloud Stream while keeping its existing Maven coordinates, Spring Cloud Gateway MVC adds WebSocket routing, and Config Server adds OAuth 2 client authorization and Google Cloud Storage backends. The train deprecates Function’s RSocket and gRPC modules alongside Stream’s reactive Kafka binder, and fixes 11 security vulnerabilities across Config, Gateway, Kubernetes, Function, Netflix Eureka, and Stream.
Alongside it, Spring AI 2.1.0 shipped with full client and server support for the stateless 2026-07-28 revision of the Model Context Protocol (spring.ai.mcp.server.protocol=MODERN), which drops session handshakes so each HTTP call can carry its own identity. It also adds a typed DecisionModel API for probabilistic yes/no, choice, and rubric calls, an append-only Session event log with turn-boundary compaction replacing ChatMemory, OpenAI /v1/responses support via spring.ai.openai.chat.api=responses, and fixes for 13 CVEs (also backported in 2.0.3).
Chrome 158 Turns Off XSLT, and Kubernetes Sets 1.40 for kube-proxy IPVS Removal -
On November 17, Chrome 158 promoted to the desktop and Android Stable channels (158.0.8127.36/.37 on desktop) and carried out its planned removal of browser XSLT for general users. Both the XSLTProcessor JavaScript API and <?xml-stylesheet?> processing instructions now stop running unless a site uses the temporary Origin Trial or an organization sets the Enterprise Policy, both of which expire in Chrome 176 next August. Pages that still transform XML or style raw RSS feeds in the browser need server-side rendering or a JavaScript XSLT library. Meanwhile on November 16, WebKit previewed spatial CSS (spatial: portal, @environment-map, and 3D position-anchor for <model> elements) behind the CSS Portals flag in Safari on visionOS 27.3.
On the cluster side, Kubernetes SIG Network announced on November 16 that kube-proxy’s ipvs backend, deprecated in 1.35, will be deleted in Kubernetes 1.40. IPVS originally beat linear iptables rules on big clusters, but the kernel IPVS API lacks batch updates, still falls back to iptables for parts of the Service spec, and cannot share scheduler state across nodes. On clusters running Kubernetes 1.36 or newer with Linux kernel 5.13 or newer, change mode: ipvs to mode: nftables in the kube-proxy ConfigMap and restart the kube-proxy pods.
NuGet Adds Staged Publishing, and the IETF Publishes EPP Over QUIC -
Until this week, running dotnet nuget push immediately put a package into NuGet’s public publishing pipeline, leaving no safe place to pre-validate an embargoed security patch or line up a multi-package release. On November 17, NuGet.org introduced staged publishing for all publishers. Using the .NET 11 SDK, upload privately with dotnet package stage push (optionally passing --group <id>) using a staging-only API key or Trusted Publishing policy. NuGet runs malware scans and signature checks while keeping the artifact invisible to search and restore until a maintainer with two-factor authentication clicks Promote on NuGet.org.
The same day, the IETF published RFC 10087, Extensible Provisioning Protocol Transport over QUIC. Domain registrars talking to top-level registries over traditional EPP-over-TCP have had to keep a separate TCP and TLS connection open for every concurrent stateful session. Using the EoQ ALPN identifier, RFC 10087 maps each stateful EPP session onto its own client-initiated bidirectional stream inside one shared QUIC connection, cutting handshake and socket overhead while keeping sequential command ordering intact.
AI & Models
Alibaba Releases Qwen4-Max With 71B Active Parameters Out of 2.4 Trillion -
On November 16, Alibaba opened Qwen4-Max on standard pay-as-you-go API pricing in Model Studio after a brief Token Plan preview. The Mixture-of-Experts model holds 2.4 trillion total parameters, matching Qwen3.8-Max in total size while activating 71 billion parameters per token, 25% fewer than its predecessor. Alibaba says it uses 35% fewer tokens at high reasoning effort on its Work360 suite and cuts average task cost by roughly 60%, with native text, image, and video inputs across a 1 million-token context window.
The model id is qwen4-max, and Model Studio endpoints accept OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages payloads alongside DashScope. International pricing starts at $1.50 per million input tokens and $4.50 per million output tokens, with implicit cache hits at $0.075 per million. Alibaba says smaller open-weight Qwen4 models will follow.
Microsoft Ships MAI Code 2, MAI Code 2 Flash, and Microsoft-Decision-1.1 -
On November 17, Microsoft brought its second-generation in-house coding models, MAI Code 2 and MAI Code 2 Flash, to GitHub Copilot across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the CLI, and the Copilot cloud agent. MAI Code 2 targets multi-file refactoring and repository debugging on paid Copilot plans, while MAI Code 2 Flash is open to every tier, including Copilot Free and Student. The same day, Copilot for JetBrains added public previews for GitHub-hosted cloud sandboxes, cross-client remote sessions, Chrome browser automation, and on-device voice dictation.
At Ignite, Microsoft also made Microsoft-Decision-1.1 generally available in Microsoft Foundry, OpenRouter, and Vercel. Rebuilt on a Gemma 4 26B A4B base with image inputs and calibrated confidence scores, it handles low-latency routing, classification, and guardrail checks rather than free-form chat.
OpenAI Measures Coding Agents Inside Its Own Lab and Ships GPT-Live-1.1 -
On November 18, OpenAI published internal telemetry on how coding agents are changing its own AI research. From March to October, autonomous completion on research tasks estimated at two to four hours of human work climbed from 23% to 46%, and four-to-eight-hour tasks rose from 13% to 35%. The gains remain uneven: agents succeeded 79% of the time when monitoring training runs, 35% when launching them, and 25% when writing inference and serving code. Priced at public API rates, the median OpenAI researcher burned $2,400 a day in agent tokens in October, reaching $23,000 a day at the 90th percentile.
On November 19, OpenAI also added GPT-Live-1.1 (gpt-live-1.1) on v1/live/sessions for full-duplex multilingual voice conversations that can listen while speaking and hand tool work to a backend model. Sessions are billed by duration at $0.05 per active minute and $0.02 per minute in WebRTC standby when audio is idle.
Security
Java4Shell Corrupts OpenJDK Heap Memory on AVX-512 Servers -
On November 17, OpenJDK and Oracle patched CVE-2026-114938, a Base64 decoder memory-corruption bug disclosed by Calif as Java4Shell (rated 8.1 under CVSS 3.1 by OpenJDK and 9.2 Critical under CVSS 4.0, with an urgent patching alert from CISA). Because HTTP Basic headers, JWTs, SAML assertions, and cookies all pass untrusted Base64 strings to the runtime before authentication finishes, an unauthenticated request can write past a heap buffer, leak adjacent memory, or run code. Calif demonstrated pre-authentication remote code execution on default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins, and PingFederate, and honeypots began logging scan traffic within 24 hours.
The bug lives in HotSpot’s hardware-accelerated Base64 intrinsic on x86-64 CPUs that support avx512vbmi, meaning Intel Xeon Ice Lake or newer and AMD Zen 4 or newer. ARM64 hosts and Java 8 and 11 are not affected by Java4Shell. Update Java 17, 21, 25, and 27 to the November 17 releases (such as OpenJDK 17.0.21.1, 21.0.13.1, and 25.0.5.1, or the matching Oracle and Corretto builds, which shipped alongside 655 total fixes in Oracle’s November Critical Patch Update). If a JVM cannot be upgraded immediately, pass -XX:-UseBASE64Intrinsics at startup and restart the process.
nginx2root Exposes Decade-Old HTTP/2 Memory Bugs in NGINX -
On November 18, F5 and nginx.org released NGINX 1.31.9 and 1.30.7 to fix ten vulnerabilities, led by five memory-corruption bugs detailed under the name nginx2root (CVSS 3.1 8.1, CVSS 4.0 9.2). The broadest is CVE-2026-123007, an integer overflow in HTTP/2 flow-control window accounting that has been in every NGINX build since 1.9.5 shipped HTTP/2 in 2015. All it needs is http2 enabled: an unauthenticated client can crash worker processes or leak addresses and corrupt the heap, with Censys counting roughly 8.8 million exposed HTTP/2 services on potentially affected versions.
Companion flaws hit specific proxy layouts, including CVE-2026-123633 (HTTP/2 with proxy_pass when worker_connections runs low) and CVE-2026-121810 (HTTP/3 with proxy_pass). Origins that sit strictly behind a reverse proxy or CDN that terminates client HTTP/2 and manages its own upstream flow control are shielded from the raw frame attacks, though any directly reachable port remains exposed. Upgrade to NGINX 1.31.9, 1.30.7, or your Linux distribution’s backported package and restart workers.
Malicious Releases Hit Microsoft’s agent-framework on PyPI and WPPConnect on npm -
On November 17, attackers used a compromised Microsoft release-pipeline client certificate to push two malicious versions (1.26.1 and 1.26.2) of Microsoft’s agent-framework package to PyPI before PyPI removed them within the hour. Both wheels dropped .pth startup hooks (agent_framework_meta.pth and af_runtime_diag.pth) into site-packages. Because Python runs .pth files automatically whenever the interpreter starts, the stealer fired even if agent_framework was never imported, sweeping /proc/*/environ across running processes, .env and registry credential files, Azure IMDS managed-identity tokens, and GitHub Actions and Azure DevOps OIDC tokens. Pin agent-framework==1.26.0, check workstations and CI runners for that one-hour window, and rotate every secret those environments could reach.
One day earlier, StepSecurity documented a supply-chain compromise in the WhatsApp bot packages @wppconnect-team/wppconnect (2.4.1 and 2.4.3) and @wppconnect/server (2.10.82 and 2.10.83), plus matching Docker Hub images. Published through the project’s own GitHub Actions trusted-publishing workflow with valid provenance, the payload ran on module import to upload environment variables, .env files, and WhatsApp session tokens to a Telegram bot, after which someone force-pushed the Git tags to hide the payload from source review.
PHP, Ruby, Valkey, pip, and CISA’s Known-Exploited Additions Round Out Patching -
Several language runtimes and infrastructure tools also shipped urgent fixes this week. On November 18, Ruby released 4.0.10, 3.4.13, and 3.3.14 for seven heap overflows, double frees, and out-of-bounds reads in Regexp and String case conversion that already have working remote code execution proofs of concept (Ruby 3.2 is end of life and unpatched). On November 19, PHP 8.5.13, 8.4.28, 8.3.37, and 8.2.36 patched critical PHP-FPM request smuggling on persistent FastCGI connections (CVE-2026-129289), session serializer and partial-write RCEs, and magic-method backtrace and timezone-cache use-after-frees, all of which are also fixed in PHP 8.6.0.
Elsewhere, Chrome 158 fixed 204 security bugs including six Critical flaws, Valkey 9.1.4, 9.0.8, 8.1.12, 8.0.13, and 7.2.16 patched a client-eviction use-after-free (CVE-2026-130822) when maxmemory-clients is enabled, and pip 26.2.2 stripped compromised eMudhra root certificates from its bundled certifi store (CVE-2026-133949). Across November 17, 18, and 20, CISA added nine actively exploited flaws to its catalog, led by a VMware vCenter authentication bypass (CVE-2026-123890, CVSS 9.8, exploited as a zero-day since September), BerriAI LiteLLM guardrail code execution (CVE-2026-40217), Perforce Puppet Enterprise deserialization (CVE-2026-115731), Apache Roller (CVE-2026-129448), and Apache Axis (CVE-2026-90436).
Funding & Industry Deals
Anthropic Goes Public on Nasdaq as ANT in a $119.5 Billion Combined Raise -
Anthropic priced its initial public offering at $83.00 per share on November 18, above its $70 to $78 marketing range, began trading on Nasdaq under the ticker ANT on November 19, and closed the offering on November 20. With underwriters exercising their full 147 million-share option, the public sale covered 1.127 billion Class A shares for $93.5 billion in gross proceeds, of which $91.6 billion went to Anthropic and the rest to selling stockholders.
At the same $83 price, Anthropic closed a concurrent $26.0 billion private placement of 313.25 million Class N shares to seven hardware and cloud suppliers: AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC. Combined gross proceeds reached $119.5 billion before underwriting fees, giving the Claude maker an unmatched war chest for frontier clusters and enterprise AI expansion.
Palo Alto Networks Buys Applied Compute for $4.5B, Isomorphic Raises $5B, and Elastic Acquires HoneyHive -
On November 18, Palo Alto Networks agreed to acquire Applied Compute in a $4.5 billion all-stock deal expected to close in its fiscal third quarter of 2027. Founded last year by former OpenAI researchers, Applied Compute builds post-training and inference software for enterprises customizing open-weight models on private and sovereign clouds; CEO Yash Patil will join as Co-Chief AI Officer. A day later, Elastic agreed to acquire HoneyHive to fold agent trace inspection and offline-to-production evaluation suites directly into Elastic Observability.
In venture rounds, DeepMind drug-discovery spinout Isomorphic Labs raised a $5 billion Series C on November 17, co-led by Greenoaks and Sequoia Capital with Andreessen Horowitz joining, to push AI-designed oncology and immunology candidates into clinical trials. Closer to everyday engineering workflows, Incredibuild took $15 million from Insight Partners on November 18 for its Islo agent-workflow platform, and Hoplite raised a $5.5 million seed on November 20 led by Stripe founders Patrick and John Collison to run parallel Claude Code and Cursor workspaces in the cloud.
Layoffs: Yelp, Crunchbase, Intel, Firaxis, and Incredibuild Cut Staff
- Yelp: On November 18, Yelp disclosed in an SEC Form 8-K that it is cutting nearly 5% of full-time employees, booking $14 million to $16 million in severance charges while redirecting cash savings into AI products.
- Crunchbase and Intel: On November 18 and 19, the San Francisco Chronicle reported that Crunchbase filed a California WARN notice eliminating 67 jobs at its San Francisco headquarters, roughly 27% of staff including several senior engineering and data analyst roles, while Intel began fresh cuts across its global IT organization alongside 57 Santa Clara roles in state notices.
- Firaxis Games: On November 17, Ars Technica reported that Take-Two’s Civilization and XCOM studio laid off about 60 employees, roughly one-third of its workforce, including studio head Heather Hazen.
- Incredibuild: On November 18, Calcalist reported that the build-acceleration vendor cut about 16 employees from a roughly 90-person staff ahead of its $15 million pivot into automated AI engineering pipelines.
The Numbers That Matter
- $119.5 billion in combined gross proceeds from Anthropic’s
ANTdebut: $93.5 billion in the public offering plus $26.0 billion from seven chip and cloud partners at $83 a share - $8 and $40 per million input and output tokens for
gpt-6.2-astra(20% below GPT-6 Astra), alongside $3 and $15 forgemini-4-neon - 55.3 milliseconds to create an in-process session and complete a warmed turn on Microsoft Copilot SDK 2.0’s Rust core, down from 5.25 seconds on the TypeScript CLI
- 30 minutes maximum runtime for on-demand AWS Lambda Event Functions, double the 15-minute limit, alongside 11,000
GET/HEADrequests per second per S3 prefix - 8.8 million internet-facing HTTP/2 services matching NGINX versions exposed to
CVE-2026-123007before upgrading to1.31.9or1.30.7 - 204 security fixes in Chrome 158 Stable, which also turns off
XSLTProcessorand<?xml-stylesheet?>by default - $2,400 median daily coding-agent spend per OpenAI researcher at API list prices ($23,000 at the 90th percentile), with agents completing 46% of two-to-four-hour research tasks solo
- $4.5 billion in stock for Palo Alto Networks to buy open-weight AI platform Applied Compute, the same week Isomorphic Labs raised a $5 billion Series C
Quick Hits
- Rust parallel frontend - November 16. Enabled by default on nightly (
2threads; override with-Zunstable-options --jobs-frontend=8), cuttingcargo checkwall time by 44% across benchmarks at 8 threads. - Kubernetes IPVS - November 16. SIG Network scheduled removal of
kube-proxy’sipvsbackend for Kubernetes 1.40; move tomode: nftableson 1.36+. - WebKit Spatial CSS - November 16. Developer preview of
spatial: portal,@environment-map, and 3Dposition-anchorfor<model>in Safari on visionOS 27.3. - Qwen4-Max - November 16.
qwen4-maxin Model Studio at $1.50 and $4.50 per million tokens internationally; 2.4T total parameters, 71B active. - WPPConnect - November 16. Compromised npm (
2.4.1,2.10.82) and Docker builds stole.envsecrets and WhatsApp sessions before Git tags were rewritten. - Azure HorizonDB and Cobalt 200 - November 17. Both reached GA at Microsoft Ignite alongside Fabric HorizonDB preview, AKS AI Manager preview, and Windows Server 2028 preview.
- Microsoft Copilot SDK 2.0 - November 17. Multi-tenant in-process Rust runtime across six languages with
mode: "empty"and resumablesessionFslogs. - Vercel eve 1.0 - November 17. Open-source production agent framework (
npx eve@latest init my-agent) with Workflow SDK 5.5 crash recovery and OpenTelemetry tracing. - GPT-6.2 Astra and Sol - November 17.
gpt-6.2-astra($8/$40) andgpt-6.2-sol($2/$10) in the OpenAI API, Codex, ChatGPT Work, and Amazon Bedrock. - MAI Code 2 - November 17.
MAI Code 2andMAI Code 2 Flasharrived in GitHub Copilot alongsideMicrosoft-Decision-1.1GA and Copilot cloud sandboxes for JetBrains. - Chrome 158 - November 17. Promoted to Stable with 204 security fixes and browser XSLT disabled by default outside Origin Trial and Enterprise Policy.
- NuGet staged publishing - November 17.
dotnet package stage pushvalidates packages privately until a 2FA maintainer promotes them on NuGet.org. - RFC 10087 - November 17. Standards Track EPP over QUIC (
EoQ) multiplexes domain-provisioning sessions as streams on one QUIC connection. - Java4Shell - November 17. OpenJDK
CVE-2026-114938heap overflow on AVX-512 VBMI x86-64 hosts; patch Java 17, 21, 25, and 27 or set-XX:-UseBASE64Intrinsics. - agent-framework PyPI - November 17. Poisoned
1.26.1and1.26.2wheels ran.pthcredential stealers on Python startup; pin1.26.0and rotate secrets. - Isomorphic Labs and Firaxis - November 17. Isomorphic Labs raised a $5B Series C; Firaxis Games laid off roughly a third of its studio (about 60 people).
- Gemini 4 Neon - November 18.
gemini-4-neonat $3 and $15 per million tokens on the paid Gemini API, plusOrchestrator (Neon)in Antigravity. - Valkey 9.2.0 - November 18. Forkless RDB snapshots and replica sync, B+ tree sorted sets, Path Hash, and
DELEX, alongside9.1.4security fixes forCVE-2026-130822. - AWS ExtendDB 1.0 - November 18. Apache 2.0 DynamoDB-compatible adapter over PostgreSQL, SQLite, Cassandra, and MongoDB (
@extenddb/devfor local runs). - GitHub CLI Event Streams - November 18.
ghv2.105.0 adds previewgh event listenso scripts and agents wait on pull request events instead of polling. - Orion Belt 1.5.0 - November 18. Relicensed from Apache 2.0 with Commons Clause to
AGPL-3.0-only. - nginx2root and Ruby - November 18. NGINX
1.31.9and1.30.7fixed HTTP/2 flow-control RCECVE-2026-123007; Ruby4.0.10,3.4.13, and3.3.14fixed seven Regexp and String memory bugs. - Palo Alto Networks and Layoffs - November 18. Palo Alto agreed to buy Applied Compute for $4.5B in stock; Yelp (~5%), Crunchbase (67), and Incredibuild (~16) disclosed cuts.
- PostgreSQL 19 - November 19.
REPACK (CONCURRENTLY),pg_plan_advice,ON CONFLICT DO SELECT,WAIT FOR LSN, and JIT off by default. - PHP 8.6.0 and 8.5.13 - November 19. Partial function application,
Io\Poll, andTime\Durationin 8.6.0, plus critical PHP-FPM and session RCE fixes across 8.2 through 8.6. - Unison 2.0.0 - November 19. Adds an opt-in LLVM 20+ JIT (
ucm --jit on), standaloneucm bundlebinaries, and anonymous records. - AWS Lambda and GPT-Live-1.1 - November 19. Lambda doubled on-demand Event Function timeouts to 30 minutes; OpenAI launched
gpt-live-1.1voice sessions, and xAI retiredgrok-build-0.1. - Elastic and HoneyHive - November 19. Elastic agreed to acquire AI evaluation and tracing startup HoneyHive.
- Spring Cloud 2026.0.0 and Spring AI 2.1.0 - November 20. Both reached GA on Spring Boot 4.2, adding stateless MCP
2026-07-28,DecisionModel, and 24 combined CVE fixes. - Amazon S3, EC2 T9g, and GKE - November 20. S3 doubled reads to 11,000
GET/HEADper second, Graviton5T9ginstances hit GA, and GKE introduced sidecarless ambient networking. - Anthropic IPO and CISA KEV - November 20. Anthropic closed its $119.5B combined IPO and private placement (
ANT); CISA added exploited VMware vCenter zero-dayCVE-2026-123890andpip 26.2.2dropped compromised eMudhra roots. - CloudWatch Omni for Azure - November 21. Amazon CloudWatch Omni added ingestion for Azure Monitor metrics, resource logs, VNet flow logs, and Application Insights traces.
PostgreSQL 19, PHP 8.6, Valkey 9.2, Spring Cloud 2026.0.0, and ExtendDB 1.0 make this one of the busiest core-stack release weeks of the year, even before counting Microsoft Ignite and Anthropic’s Nasdaq debut. On the maintenance side of the ledger, OpenJDK’s AVX-512 Base64 decoder, NGINX’s HTTP/2 flow control, PHP, Ruby, and any runner that touched PyPI on Tuesday afternoon are the places worth checking first.