Habitat H4 · 233 sightings
Running AI: infrastructure, deployment, cost, security, and operations. Every sighting here, newest first, with its specimen number and wildness rating.
Microsoft says servers failing automated health checks cut Exchange Online access and broke Copilot. It found no sign of an attack and promises a report.
A race in mod_http2 lets an unauthenticated client corrupt httpd memory and run code. Upgrade to 2.4.71 or turn off h2 and h2c until you can.
Kubernetes from Zero · Foundations 2 of 12 · Lesson 2
Name every control-plane and node component, then follow one kubectl apply step by step until a container runs.
Daybreak Red Team sends a team of agents at your web apps and is in early access. Defense Factory turns findings into patches that engineers approve.
CoreWeave's managed PostgreSQL runs on CPU nodes in your CKS cluster beside your GPUs, with HA, point-in-time recovery and pgvector. Built on CloudNativePG.
A lab demo, not a Claude flaw: with a hacked machine and a stolen API key, Darktrace steered Claude Code through memory stores and pulled out a 1 GB file.
OpenShell, NVIDIA's runtime that enforces what AI agents may touch, is now a CNCF Sandbox project. It moves to a neutral ai-openshell GitHub org.
DuckDB's team explains why storing data by column makes it fast. We tried the ideas on a small LLM eval log, with short queries you can run in a minute.
A critical use-after-free in MongoDB Server's replication code is reachable without credentials. Fixes are out for 4.4 through 9.0. Patch every member.
Since 10 November, stolen GitHub logins have opened probe PRs in 60,000+ repos, then pushed a stealer that grabs CI secrets and deletes its own run logs.
Kubernetes from Zero · Foundations 1 of 12 · Lesson 1
Why containers led to Kubernetes, what desired state means, and your first cluster answering on your laptop with kind.
AIR Security says Anthropic's docx, pptx and xlsx skills load a LibreOffice shim from a fixed temp path without checks. Anthropic calls it out of scope.
Siemens patched a 9.2 CVSS flaw where Mendix Studio Pro's MCP server was reachable from the network. A fast-agent advisory shows the same default can mean RCE.
Vates says attackers are exploiting a critical Xen Orchestra login bypass that can lead to root on XCP-ng hosts. Update to 6.9.2 or 6.10.1 now.
Mittwald says attackers who used its MCP server to reach 56 projects are back. Its website was hit and it shut its customer panels and SSH on 12 Nov.
AI agents wrote about 6.5 million lines of Lean proving that libsignal's compiled AArch64 code keeps Signal's security promises. Humans approved the specs.
Managed Postgres is GA on AWS, with backup storage and egress now billed. Xata's copy-on-write branching and scale-to-zero are coming to it; no date given.
Nitro Sandboxes are GA in all commercial AWS Regions at no extra charge. AWS pitches them for running LLM-generated code and agents on shared EC2 instances.
strutil-core 1.0.1 stole credentials on install and planted MCP servers, hooks and AGENTS.md files so coding agents would rerun it. Pulled within hours.
Two more eighth-gen TPU systems, a TPU agent suite for coding tools in preview, GKE dynamic slicing GA and an Apache-2.0 C++ serving runtime. No prices yet.
Official Hermes plugin threat-models a local repo, reviews it with parallel subagents, can test findings and exports SARIF. Your model sees the code.
Microsoft's azd Foundry extension is GA at 1.0: hosted agents live in azure.yaml, test locally, deploy with azd up and gate CI on evals via GitHub Actions.
Botpress says a staff API credential read bot configs in 16 workspaces on 21 Oct, 07:44 to 08:23 UTC, and says to treat those stored credentials as exposed.
On 5 November ASD's ACSC published two papers on open-weight models: inventory them, check hashes and signatures, and approve risky actions outside the model.
CVE-2026-116635: before 0.2.1, a crafted project name let an influenced agent read JSON outside its folder, or overwrite it with writes on. Upgrade.
Scope asked 20 agent setups to sign up for 18 hosted databases and use them. Prisma Postgres led at 88.3%; seven products had no successes.
HashiCorp's GA Terraform policy agent turns Sentinel and OPA rules into Terraform policy with tests and reports, using Copilot, Bedrock or LiteLLM.
Google Agent Substrate 1.0 is GA on GKE; Substrate, gVisor and OpenRL enter CNCF Sandbox. Vendor claims sub-500ms restore and ~10x density.
Public repos lose pull_request_target by default on 2 December, with brownouts on 9 and 16 November. Actions lockfiles that pin action commits are GA.
Dropbox files an 8-K for a ~$350M cash deal for Tigris, expected to close in Q4 2026, and launches S3-compatible DBX Agent Cloud.
AlloyDB Express gives developers and coding agents a 1 or 2 vCPU Postgres 18 database in seconds, scales compute to zero and has a 12-month free tier.
Google says its Project Suncatcher prototype ran Gemma 4 on a data center class TPU in orbit. Next: radiation and heat tests, then two satellites in 2027.
Render's agent sandboxes run as long as needed, suspend when idle and restore memory from snapshots. Oregon only, with key brokering in early access.
Binance's initial report: a tampered Homebrew install of an AI coding tool, a malicious interview repo, then an internal AI tool used to reach engineers.
A GitHub issue quotes Postman 12.33.4's own log installing Claude Code and Codex plugins, a startup hook and an MCP server. How to check your machine.
Playwright 1.64.1 still ships Chromium 156.0.8078.4, older than Chrome's 52-fix update, and runs it unsandboxed by default. What agent builders can do.
Paige and the renamed PD Reliability Platform are GA as of 3 November. Autonomous fixes are opt-in and marked coming soon; PD Labs previews are unsupported.
CVE-2026-112816 is rated CVSS 10. Canonical shipped fixed LXD and MicroCloud snaps on 29 October. Refresh, then check which servers listen on the network.
KMS now offers ML-KEM keys, Aurora PostgreSQL 18 lets you pick post-quantum TLS groups, and a new CloudFront header shows which viewers use PQ key exchange.
Grafana 13.3 adds cross-dashboard variables, Notebooks and a UI refresh in public preview. Here are the toggles, the plugin impact and the one breaking change.
Azure WAF's HTTP DDoS ruleset for Front Door Premium is GA. It learns each profile's traffic and starts in Log mode. Its two geo rules are still in preview.
EKS Auto Mode now turns Gateway API resources into AWS load balancers: ALB for HTTP and gRPC routes, NLB for TCP and UDP. Ingress keeps working alongside.
WorkOS found unauthorized access to a sales tool on 30 October. CRM contacts and customer call and Slack excerpts were exposed; it says production was not.
Microsoft patches the AKS app routing add-on's NGINX ingress only through November 2026. Its Gateway API successor gets a new IP, so plan a DNS cutover.
EKS clusters created since 2018 have 10-year CAs nearing rotation. AWS rotates them, but nodes you manage and kubeconfigs must trust the new CA first.
Trigger.dev blocked the exploit and emailed affected owners. It has not said how many orgs, how, or for how long. Rotate what you stored there.
Azure Functions ends support for the in-process .NET model on 10 November, the same day as .NET 8, .NET 9 and PowerShell 7.4. Microsoft points to .NET 10.
Tencent says agent-created database instances grew over 80x in a year, with 1 million active, and added branching and faster serverless scaling.
GitHub's hosted runners are thrown away after each job. Your own runner is not, and GitHub's own guidance says it should almost never serve a public repository, because any pull request can run code on it.
On Linux, load average includes tasks in uninterruptible sleep, waiting on disk or some locks. A load five times the core count can mean a machine stalled on memory or I/O, with CPUs that are not even full.
docker run -p 8080:80 listens on all host addresses, IPv4 and IPv6, and Docker's own docs say its rules bypass ufw. Writing 127.0.0.1 in front of the port is the difference between a local service and a public one.
Files under /dev/shm and other tmpfs mounts live in RAM and swap. A df on the main disk never counts them, so a forgotten build folder there can starve a machine while every disk alarm stays quiet.
CVE-2026-116366 is rated 9.5 under CVSS 4.0. Fixed in four ClickHouse lines; Cloud is patched. Langfuse pinned its compose file. What to upgrade.
Cloudflare logged 22,000 handshakes using eMudhra's weak keys and found bots, old browsers and WebViews still accept them. Check your stack.
A forwarded SSH agent can be used by anyone able to bypass permissions on its socket. ProxyJump reaches the inner server over a TCP forward with no agent needed.
Northflank declared a security incident on 30 October and told users to rotate all secrets. It relays user reports that stolen AI keys were used.
Radware found one-day campaign items in repos such as Trivy, LangExtract and MLflow. Two runs show the injected command executed in GitHub Actions.
LanceDB's Python SDK no longer runs remote model code or calls endpoints named in a table's metadata unless your process approves them.
Korea's FuriosaAI closed a $620M Series D at a $2.9B valuation. It plans 20,000 RNGD cards this year and a 2nm Gen 3 chip with Broadcom.
A Cosmos DB sample gives each coding task its own emulator, worktree and seeded fixtures, so parallel agents stop overwriting each other's test data.
pg_pinecone keeps pgvector tables and queries but adds a native engine with filtered search and no network calls, plus an optional Pinecone cloud mode.
Akamai Inference (12 open-weight models, serverless or dedicated) and Akamai Agent Sandbox, built on the open-source Spin framework, are both in beta.
Long prompts to GLM-5.3 on Workers AI could return text from other users' requests from 27 Aug to 9 Sep. Cloudflare and SGLang explain the bug and the fix.
CVE-2026-115812 (CVSS 9.8) lets a crafted model archive write files outside its folder. MMS is end of life, so AWS says migrate. No fix is planned.
Oracle's Deep NL2SQL agent grounds questions in a business knowledge graph, asks when unsure, and reruns vetted gold SQL before any knowledge change ships.
Forever Security found an agent instructions file in a customer's Downloads folder. Its tests hijacked Codex and Claude Code. Never start an agent there.
The official MCP debugging tool published nine security advisories on 21 October. Two critical bugs let a remote MCP server run local commands. Fixes are out.
Open Linux modules lift GeForce FP16 and BF16 tensor issue rates toward 2x on RTX 30 and 40 cards. Owners report 14 to 36% faster LLM prefill, not decode.
Devin Vulnerability Fixes is an MIT set of SKILL.md skills: inventory, an approval gate, verified versions, tests in the new image, one PR per repo.
AWS Continuum for code vulnerabilities is generally available in two US regions. It sandboxes exploits, ranks by business impact and opens fix PRs you approve.
Before 0.40.0, LanceDB's Cloud client put table names into REST URLs unencoded, so a name with ../ could reach another table or tenant. Upgrade.
In preview for GitHub Enterprise Cloud, it blocks malware, risky versions, new install scripts and too-fresh npm packages before agents or CI install them.
Postgres committer Robert Haas says LLM bug reports pushed back-patch commits up over 60% this year and pulled review time from the PostgreSQL 19 release.
Pillow 13.0.0 lists 38 security fixes, two rated critical, including EPS argument injection and a JPEG bug Anthropic found with Claude. Upgrade and rebuild.
IBM Research's vLLM-Hook MIA plugin captures attention and activations, and steers them, inside vLLM 0.29 with CUDA graphs left on. Overhead is a few percent.
Anthropic's red team and Theorem used Claude to write Rocq proofs for 123 open-source libraries. The proofs carry real assumptions and cost many tokens.
Railway paused deploys, killed API tokens and reset bucket and GitHub credentials. It says to roll every environment variable and check AI provider bills.
Microsoft says a China-aligned group used ARTEX agents after the ConnectWise Cloud Backup breach, trying 2,250+ Microsoft 365 tenants. What to check.
DuckDB 2.0.0 is out with Quack client-server, async I/O for S3, VARIANT, agent mode and triggers. Support runs to October 2027. Test before you upgrade.
PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical RCE bugs reported with AI help. AWS logged an attempt on 17 October, four days before disclosure.
Poisoned LiteLLM 1.107.0 wheels and images stole cloud, SSH and AI keys, plus provider keys stored in the proxy database. What happened and what to do.
EAGLE-4 adapts its draft tree to batch size and hardware. It needs patched engine builds, and commercial use stops being free above $20M in revenue.
A Symantec honeypot captured an offensive agent's full prompt: 74 tools, up to 100,000 unattended steps, run on stolen inference. How to close the door.
PingCAP says stolen service-account credentials let an attacker copy TiDB Cloud records and source code from 9 Sep to 7 Oct. Rotate management API keys now.
In 90 of 6,637 malicious agent skills, VirusTotal found the payload hidden where reviewers don't see it but agents still read it. Here is how to check.
TorchTPU, built with Meta's PyTorch team, is open source. Google adds TPU backends for vLLM, SGLang, Ray and VeRL, and device_map="tpu" in Transformers.
Modal's first Vera CPU server matched x86 on quiet single tasks but kept its pace as sandboxes piled up, doubling throughput in Modal's own early benchmark.
doubleAI's Maxwell optimizes programs while they run and says it formally proves each rewrite equivalent. It launched with a $200M Series A disclosure.
llm-d 1.0, a CNCF Sandbox project started by Red Hat, IBM, Google Cloud, CoreWeave and NVIDIA, adds flow control, router HA and DisaggregatedSet.
DeepInfer is Apache-2.0 and runs V4.1-Flash on NVIDIA SM100 or Huawei Ascend 950. This release is offline inference only, without KV cache or load balancing.
AWS's TorchNeuron makes Trainium2 and Trainium3 a native PyTorch device, so .to('neuron') and torch.compile work. Apache-2.0 public beta opens 20 October.
OpenCode says a breach hit 1.4% of OpenCode Console users with linked providers. If you got its email, revoke those keys at each provider.
Anthropic says an internal model traced the NIST P-256 seed to 16 August 1997. Filippo Valsorda says he reproduced it. Users need not switch curves.
CERT-In says unnamed AI providers disclosed that agents reached Indian organisations' live systems and read personal and confidential records.
On 10 February 2027 Let's Encrypt defaults to 64-day certificates unless a subscriber picks 45 or 6 days. It expects the last 90-day one to expire 11 May 2027.
Cloudflare's 15 October post says a July pilot on about 1% of its network fed discovery signals to OpenAI's search crawler. The page count is Cloudflare's.
On the Q3 call, Moynihan tied a couple of hundred million plus to AML and Mythos work, and said 1,000 consultants were added.
Anthropic says daily merges rose from around 2,500 in May to over 30,000 in September, and that seven people built Scion in five months.
HiddenLayer says a Jev checkout agent gave a planted Confirm Price button 91 percent. The Register says TypeSafe did not immediately reply.
Vercel's 14 October post counts 1,285 sandbox-challenge reports and $242,000 paid, with 57 valid findings. A Linux kernel defect is the critical one.
Noma says an HTML file over 256 KB in Claude Code Desktop's preview pane could read local files. The CVSS 8.7 score and the 1.30096.1 fix are Noma's account.
Tiger Data's 13 October post introduces pg_vectorsearch, an alpha Postgres extension with a PRISM index. The post says pgvectorscale moves to maintenance.
Cloudflare's 13 October changelog says the API MCP server posts Workers Issues to a client webhook. Issues is an open beta and free during the beta.
NVIDIA's 13 October post says Run:ai is open source under DSX. The platform and its component repositories are under Apache 2.0.
At OCP, Intel named Crescent Island the GPU D-series. GPU D270 is a 350 W air-cooled PCIe card. Specs are Intel's. No price or ship date.
Split incident duties into named roles with explicit handover rules, so the people fixing the problem are not also writing updates and keeping the timeline.
Add summary and runbook_url annotations to Prometheus alerting rules, render them in Alertmanager templates, and test them so the on-call link never goes missing.
Use an Istio VirtualService to copy live requests to a candidate model deployment, discard its responses, and compare its errors and latency with the stable version.
Liveness decides whether Kubernetes restarts a container. Readiness decides whether it gets traffic. Wiring one health check to both turns a slow dependency into a restart storm.
Linux records how much CPU time a process has used, never a percentage. Every percent figure is a difference over an interval, which is why ps and top disagree and why a tool can print a confident zero.
MemorySwapMax=0 on a systemd unit stops that one service from being swapped out while the rest of the machine still can be. How to set it, read it back and undo it, and what it costs.
A process list sorted by resident memory can hide the largest holder, because memory pushed to swap no longer counts as resident. Where Linux records the swapped part, and a one-line way to rank by both.
Two machines' logs that don't line up can mean one clock isn't synchronised, or that time zones got mixed. How to check sync on Linux, and which clock to use for timestamps and for durations.
logrotate can rotate a log by renaming it or by copying and truncating it. One can lose lines, the other needs the program to reopen its file. How to tell which your service needs.
The same line in an environment file can give three different values in Docker Compose, a systemd unit and a shell. Three documented differences: comments, backslashes and spaces.
Both run a job on a schedule. They differ in what happens when the machine was off, when the last run is still going, where the output ends up, and who hears about a failure.
At the OCP summit in San Jose, NVIDIA and AMD each described a rack for AI agents, apart from GPU inference. Both are designs and plans, not shipping tests.
Scott Guthrie says Azure Maia 300, Microsoft's third in-house inference chip, begins production deployments in 2027 and is meant to be the company's main inference system. The efficiency figures are Microsoft's.
Two 21 September commits added a loader and a root-level preinstall key to Zod's git repo and were reverted that day. The npm tarballs for zod and @zod/mini 4.6.6 do not contain them.
JFrog rates CVE-2026-105192 at 9.8 for pickle on LMCache's multiprocess ZeroMQ socket when that socket is routable. No release removes the path. Pull request 5644 is open and unmerged.
MIG cuts a supported NVIDIA GPU into up to seven instances with their own memory, cache and compute. How to create them and pick a profile for a model server.
The release team is keeping PostgreSQL 19 RC1 on 15 October 2026, adding an RC2 on 12 November, and moving general availability from 29 October to 19 November. The roadmap still said October.
An inactive PostgreSQL replication consumer can retain write-ahead log files. Learn how to inspect slot state and choose between unbounded retention and a finite recovery window.
Pipelining cuts Redis request round trips, while MULTI/EXEC controls whether a group runs without another client interleaving. Choose by the state guarantee each operation needs.
Twilio outbound media expects base64-encoded raw μ-law audio at 8 kHz. A voice model’s PCM or WAV output needs explicit conversion before it reaches the call.
One missing IP fragment can erase a whole UDP datagram. Path MTU discovery and application-sized messages help voice systems avoid that failure mode.
A Linux TCP connection can look healthy while an AI request stalls. Keepalive, user timeout, and an application deadline answer different questions.
HTTP/2 interleaves independent request streams, yet one missing TCP segment can delay them together. QUIC changes stream delivery, with limits of its own.
A trusted TLS certificate can belong to the wrong service. Follow the reference hostname into subjectAltName and reject mismatches without Common Name fallback.
A trusted certificate can coexist with an unexpected HTTP protocol. Learn how ALPN selects h2 or http/1.1 and what to verify at a voice or AI API edge.
OpenTelemetry resource attributes identify the emitting service and environment. Keep request details on spans so voice application traces group by a stable service identity.
A burst of voice worker errors can exceed journald’s per-service limit. Learn how its interval, burst and disk rules affect what an incident log can prove.
What systemd waits for with Type=notify, when an application should send READY=1, and why short-lived helpers can lose notifications.
How systemd restart policy meets start-rate limits, and why clearing a failed counter is a separate step from fixing a service.
Why a removed log can still consume disk space, and how to identify the process holding it before taking action.
A Lua script can keep a Redis read-and-update decision together, but every millisecond of execution holds up other server activity. Bound the work before using EVAL.
Redis memory pressure can evict a cache key or reject a write. Choose the policy by deciding which data may disappear and how the application handles a full instance.
Marco Bambini's sqlite-multiwriter, a VFS extension created 8 October 2026, claims concurrent writers on one SQLite file. The 5.7x and 3.0x figures are the project's own runs.
Barracuda describes a phishing email aimed at the person and at the AI assistant that summarizes the inbox. The research gives no count of how widespread the campaign is.
A restarted voice worker can make a cumulative request counter fall. Use Prometheus rate on each series before summing replicas, and keep gauges for values that fall normally.
SQLite 3.54.0 is dated 9 October 2026, ahead of the 15 October date still on its draft changelog. The release overhauls the shell, drops Windows XP, and adds limits for untrusted SQL.
DuckDB's v2.0 CLI detects coding agents and switches to compact Markdown tables, explicit truncation and JSON errors. DuckDB says agent output fell 59%, about 0.5% of total input tokens.
Committer Peter Eisentraut says 8 to 10 significant features were reverted from PostgreSQL 19 in beta. He points to LLM-assisted review finding defects faster and security work cutting review time.
NVIDIA's bulletin rates CVE-2026-47483 at 8.2: unauthenticated pprof requests can exhaust DCGM Exporter. Lava found about 2,100 hosts exposing exporter metrics. NVIDIA lists 4.8.2 as the fix.
Kaarel Moppel's self-funded cloud runs put PostgreSQL 19 Beta 3 and Beta 4 about 3 percent ahead of 18.6 on his OLTP tests, with pgbench inserts about 25 to 30 percent faster. It is one person's beta benchmark.
incident.io's 6 October post says each on-call tick ran in a savepoint, so SKIP LOCKED workers piled up on MultiXact locks. A lease, no subtransactions, and one claimer per pod preceded a 3.5 times gain.
PlanetScale says TIN 1.0.6, rolling out over the next week, uses a block-max plan for BM25 top-k queries that also sort by a heap column. Stemming for 18 languages arrived in 1.0.4 and needs a rebuild.
Goodfire says a probe plus a judge reached about 93 percent recall on harmful Kimi K3 sessions, at roughly 50 times lower cost than judging every turn. TechCrunch says Baseten customers can enable the monitors.
A voice pipeline can fail to create another tiny segment while capacity still looks available. Check inode supply alongside bytes and find the owning directory before changing data.
Creating a DNS record for a new voice endpoint does not clear earlier negative answers. Learn how NXDOMAIN, NODATA and SOA-derived lifetimes affect rollout checks.
A voice service can receive a different setting from the one Compose used to build its configuration. Learn how defaults, host inputs and escaped dollar signs cross that boundary.
A bind mount overlays its container target, hiding files packaged there. Read-only protects the host source from container writes but does not preserve obscured image files.
Kubernetes refreshes ordinary Secret volume files eventually, but a subPath mount stays on its old content. Credential rotation also needs an application reload or a Pod rollout.
BuildKit secret mounts make credentials available only to the build instruction that needs them. ARG, ENV and copied files can disclose values through image metadata, layers or cache.
EndpointSlice conditions distinguish an endpoint that is serving from one accepting normal new traffic. Use ready, serving, and terminating to design a drain policy for long voice or AI sessions.
Neon says the lakebase_tokenizer extension, packaged with Lakebase Search, stores synonyms and stop words in SQL tables. The post does not name a new plan or region.
PGX says pgx-bm25 1.0 is a PostgreSQL-licensed BM25 index for PostgreSQL 17 and 18. The index lives in the server's own pages, so it gets WAL, crash recovery, and VACUUM.
ClickHouse's 5 October post calls executable UDFs generally available on AWS, GCP, and Azure, with a Native runtime beside Python 3.11. Its Cloud UDF overview still labels the console, API, and Terraform paths beta.
Averaging latency percentiles across replicas can misstate an endpoint’s performance. Histograms let you combine observations before calculating p95.
Memory requests place a Pod. Limits constrain its containers. Learn how to tell a container out-of-memory kill from node-pressure eviction and size inference workloads accordingly.
Dependabot grouping rules decide which security fixes arrive together. Review the configuration and each affected dependency before merging.
OpenTelemetry baggage can leave an AI service in outbound HTTP headers. Check what a model API receives and limit propagation at that boundary.
A replica count alone does not spread Pods across failure zones. Kubernetes topology spread constraints let you define the placement you need and the tradeoff when a zone is unavailable.
An incomplete S3 multipart upload can leave billable parts behind even when no finished object appears. Here is how to find and clear them.
A Collector can accept spans while its exporter falls behind. Its internal metrics show where delivery slows, where data is refused, and when loss becomes likely.
An open read transaction can keep SQLite’s write-ahead log growing even while automatic checkpoints run. Here is why, and how to find the cause.
A Kubernetes scratch volume can hold files a model server prepares at startup. Its lifetime makes it a poor home for the only copy of a model or a result.
An ephemeral container gives you a way to inspect a running Kubernetes pod when its application image lacks a shell or diagnostic tools.
A trace can show which model calls and tool runs produced an agent response, where the time went, and how many tokens each call used.
Use a Kubernetes init container and a shared volume to put model files in place before an inference server starts. Plan for retries, storage use, and model loading.
A 429 can tell an agent when to retry and which limit it reached. The response body also matters: some 429s will not clear with backoff.
A PodDisruptionBudget can limit how many inference pods a node drain evicts at once. Its allowed-disruptions count shows when the workload has room for another eviction.
ParadeDB tagged v0.26.0 on 3 October 2026. The notes include vector pushdown for RRF queries and ORDER BY that can push a vector distance plus a tie-break column. The pull requests merged in July.
Kubernetes can give urgent Pods earlier scheduling attempts without removing running work. Non-preempting priority still depends on suitable capacity.
GitHub's 30 September 2026 changelog adds an opt-in Allow npm dist-tag permission on trusted publishing, so an OIDC workflow can move latest, next, or beta without a long-lived token. It defaults to off.
A response can be slow to start or slow to continue. vLLM exposes separate metrics for those delays, making each easier to investigate.
Cloudflare's 1 October 2026 post adds ML-KEM and ML-DSA to Workers Web Crypto, off until you set webcrypto_modern_algorithms. ML-KEM-512 is absent. SHA-3 and HPKE are not in this change.
Cloudflare's 1 October 2026 post introduces Workers KV Instant, a private beta that replicates keys through Quicksilver. Reads are $0.20 per million, and a namespace tops out at 1 MB.
Cloudflare's 2 October 2026 post opens Cloudflare Traces in beta: one OpenTelemetry timeline for supported security, cache, routing, Worker, and origin steps, exported over OTLP. New ingest pricing starts 1 December.
Head sampling controls trace volume early. Tail sampling can retain runs that later fail or become slow, if the pipeline can hold and route their spans.
Cloudflare's 1 October 2026 post opens a waitlist for a managed Cloudflare OS deployment. Self-hosting remains. The same post says the workspace can now use a GitHub repo and more of Google Workspace.
Cloudflare's 2 October 2026 post adds an OHTTP Gateway in closed beta, as a paid add-on with a waitlist. The older Privacy Gateway relay is renamed OHTTP Relay. The gateway refuses Cloudflare-proxied clients.
Cloudflare's 1 October 2026 post launches K2 in public beta: an ordered event log stored as files in R2, with competing consumers or fan-out. Cloudflare says produce latency is about 1 second at p99.
Cloudflare's 1 October 2026 post says the data platform from Birthday Week 2025 is now Basin, and that Pipelines, Catalog, and SQL are generally available on Apache Iceberg in R2.
CPU requests guide Pod placement. CPU limits cap a running container's CPU time.
Inference requests can wait while CPU utilization stays below its scaling target. Measure pending work and queue delay to decide when more workers are useful.
A multistage Docker build can compile an application in one stage and copy the finished binary into a smaller final image.
Aviram explains why separate credentials, limited permissions and protected backups matter when AI agents work with real infrastructure.
A sudden inference surge meets several gates before new Pods serve traffic. Follow the autoscaler's measurement loop, replica rules, Node capacity, and readiness.
Multipart uploads and some encryption modes produce ETags that cannot be compared with a file's MD5. Use an explicit S3 checksum when verifying object data.
A lasting agent operation has a start and an end. Named, timestamped events show when approvals, state changes, and failures occurred along the way.
A user ID on a Prometheus metric creates series across other label dimensions. Here is how that affects storage and queries, and how to keep useful operational detail.
A build attestation can connect a downloaded binary to its source repository and build workflow. Verification still leaves the decision to trust that code with you.
Creating a Kubernetes NetworkPolicy does not mean its rules are already enforced. Learn what can happen during the handoff and how to check the result.
A Deployment can exist even when its Pods cannot be created. Check namespace quota usage and resource declarations before investigating node placement.
Device plugins advertise GPUs as countable resources. That changes how Pods request them and how many workloads a node can schedule.
A Kubernetes progress deadline reports a stalled rollout. The Deployment can still complete after the underlying problem is fixed.
A GitHub Actions job can use OpenID Connect to obtain short-lived AWS credentials. The role trust policy decides which job may deploy.
Docker and Kubernetes ask a container to stop with SIGTERM and then kill it. Why some services never hear the request, and how to make sure yours finishes its writes before the grace period runs out.
A read replica lags its primary and is read-only. Code that treats it as the same database can read stale data, or find its queries cancelled mid-flight.
Cloudflare has launched a Web Search API through AI Gateway with Ceramic.ai, Exa and Linkup. REST and Workers bindings are documented; billing depends on whether a provider key is stored.
A tag like :latest or :2.1 is a name that can be moved to a different image at any time. A digest cannot. What that means for what you run, and how to pin it.
An open port and a running process say little about whether a service actually answers. How to write a health check that asks the real question, in Docker and Kubernetes.
API keys rarely leak through the vault. They leak through a command line other users can read, a URL that gets logged, and a conversation transcript that gets archived.
Agent logs are tempting to fill with everything: prompts, tool calls, environment, full requests. A short list of what must never land in them, and why scrubbing has to happen before the write.
ClickHouse (Oct 2, 2026): Managed Postgres live wal-g backups use WALG_DIRECT_IO plus stripe-sized RAID0 NVMe reads to skip page-cache eviction. Vendor benches only; incremental still prototyping.
A full disk breaks everything at once. Where the space usually hides on a small Linux server, and the warnings worth setting before the last few percent go.
Supabase (Oct 2, 2026) is acquiring Turso. Supabase continues Postgres; Turso continues SQLite. Glauber Costa joins with Pekka Enberg and the Turso team to lead agentic infrastructure. No deal price is disclosed.
When an AI agent is not sure a file is safe to delete, the right move is a recoverable one: move it to a holding folder with a note, and let the owner decide.
A backup file you have never restored is a hope, not a backup. A short restore drill for small servers, and the two database mistakes it catches most often.
Several AI coding sessions sharing one server will collide over ports, checkouts, containers and disk. Four small habits that keep them out of each other's way.
JVN#22475874 (2026-09-29): seven Pgpool-II CVEs; lead CVE-2026-92867 authenticated OOB write (CVSS 4.0 8.7 / 3.0 8.8)—process termination and arbitrary code execution. Fixed in 4.7.3, 4.6.8, 4.5.13, 4.4.18, 4.3.21.
Radim Marek (boringSQL, Sep 28) benches PostgreSQL 19beta4 core REPACK (CONCURRENTLY) vs pg_repack/pg_squeeze: WAL, disk, cluster-wide VACUUM hold-back, ~105M change ceiling, and MVCC caveats—author benches only.
Redis advisory CVE-2026-81934 (Aug 28, 2026): authenticated TLS pending-data use-after-free; public CVSS revised 9.8→7.5 High. Upgrade OSS/Software builds; Cloud Essentials patched, Pro underway.
Redis operator 7.22.2-45 (Sep 2026) fixes a high-severity Active-Active bug: GitOps/config updates could send an empty password—DB stayed healthy while accepting unauthenticated clients. No CVE for this bug.
Google Cloud AlloyDB “PostgreSQL for agents” (docs updated Sep 24, 2026) is Preview only: agents query live transactional data with isolation from core systems and can scale compute to zero when idle.
ClickHouse (Sep 29, 2026): Fabric workload in public preview; OneLake Iceberg read GA via Table APIs; Iceberg write in public preview with credential vending; Azure BYOC on the Microsoft Marketplace.
ClickHouse’s chDB Durable Layer (Sep 28, 2026): local MergeTree plus object storage via flush()/checkpoint(). chDB 4.4+ with chdb[durable]; chdb-core 26.7.3; single-writer. npm [email protected] differs from the Python package.
MySQL and MariaDB now document native vector features, PostgreSQL can use pgvector, and SQLite can add search through extensions. Here is how to choose by deployment shape and retrieval needs.
PgBouncer 1.26.0 (Sep 23, 2026) fixes CVE-2026-19888, CVE-2026-6668, and CVE-2026-6669 (DoS only: crash, infinite loop, unbounded login work; no RCE or CVSS). Also: pool_idle_timeout and query_wait_timeout; -R removed.
PlanetScale’s hot-shards post (Sep 28, 2026) walks multi-tenant AI SaaS through vertical scale, whale isolation, and online table-level Reshard on Neki—without taking the app offline. The post gives no Neki GA date beyond “since the launch of Neki.” The Slack Vitess analogy is third-party; Slack does not run on Neki. “Fastest cloud Postgres” is PlanetScale’s own claim.
MongoDB launched Atlas Agent Engine in public preview (press Sep 29, 2026): unified execution, memory, and governance on Atlas with Voyage AI retrieval. Not GA, and no GA date is given. Preview pricing subject to change. Not a substitute for Atlas Vector Search; separate from the Voyage/Atlas $rerank deep dives. Framework names are used in Atlas’s framing only.
Redis 8.4 adds optional CLAIM min-idle-time on XREADGROUP—claim idle PEL entries then read new ones in one command (shared COUNT). Idle ms + delivery count come back when CLAIM is used; XACK still required. ~22.5× vs XAUTOCLAIM is Redis-reported on their stress setup.
Qdrant’s Constella research preview (Sep 29, 2026): index docs once with Stella (400M English), then query with Zero, Nano, or Stella against the same collection—no re-embedding. Research preview, not GA. Encode-speed and nDCG figures are vendor-reported.
DiskANN approximate vector indexes are GA in Azure SQL Database, Azure SQL Managed Instance (Always-up-to-date), and Fabric SQL (Hoffman Sep 29; Thota Sep 28). Vectors sit next to operational rows under the same T-SQL/security model—no separate vector store. SQL Server 2025 / MI on SQL Server 2025 policy stay preview.
Databricks GA’d Lakebase Search on AWS/Azure (blog 2026-09-28): lakebase_vector (ANN) + lakebase_text (BM25) in the same serverless Postgres OLTP—hybrid via RRF, no separate search cluster. VectorDBBench/Conexiom figures are vendor-reported.
AWS GA’d direct Iceberg/Parquet lake queries from Aurora PostgreSQL (2026-09-30) via embedded DuckDB and aurora_analytics—no ETL. Join live ops rows (incl. uncommitted) with the lake. PG 17.11+/18.6+; no extra feature fee (compute + S3).
AWS S3 Vectors (Sep 30, 2026) adds ENHANCED index mode that filters metadata before similarity search, plus $startsWith. AWS claims up to 5× more matching vectors on highly selective filters—vendor figure, not house-verified. No extra cost; CLASSIC until you upgrade.
pgvector, JSONB, full-text search and row security can support an AI app in one PostgreSQL database, with practical limits to measure before you scale.
A nearly full disk on a machine running AI review sessions: what was using the space, the explanation I checked first, and the rule that would have prevented it.
A practical guide to using Redis for response caches, request limits, agent queues, session memory and vector retrieval, with the correctness and licensing trade-offs to check first.
AWS's Rust SDK release of September 25 carries the Well-Architected Agent, a generative AI service that reviews an AWS environment and ranks fixes for cost, security, performance and resilience.
Microsoft's September 23 Surface launch never says Copilot+ PC, and a Surface executive told Windows Central the new machines are not called that, though Microsoft has announced no formal retirement.
