JadidJust-in AI Developments, Incidents & Deadlines
October 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

The agents are outgrowing the tools built for people: Anthropic's merge volume broke Git, OpenAI's always-on assistant burned 300K tokens without a prompt, and npm's publishing trust model turned out to hinge on a repo name anyone could reclaim.

23 stories · transcript

Meanwhile the vendors keep pulling the floor out from under older models, so today's deadline is as much a migration story as a news story.

TLDR

  • OpenAI deprecates GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini, with API shutdown on Jan 14, 2027; fine-tunes go with them (Oct 14) Pricing
  • SAP Cloud SDK for Python has a CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 10.0 flaw that leaks OAuth2 client secrets, and Docker Hub closed a two-year token-takeover bug (CVSS 9.4) (Oct 12–13) Security
  • npm trusted publishing could be hijacked by re-registering a freed GitHub repo name; 1,500+ packages were exposed (Oct 13) Security
  • Anthropic replaced Git with an in-house system, Scion, after merges grew from ~2,500 to 30,000+ per day (Oct 13) Production
  • Claude Haiku 5.5 cuts small-model pricing to $0.10/$0.50 per million tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ (Oct 7, first coverage) Pricing

Upcoming Deadlines

Date What Who it hits Do this
Wed Oct 14 (today) GPT-5.5 retired from ChatGPT, ChatGPT Work and Codex (API unaffected) Anything pinning gpt-5.5 in Codex or ChatGPT Work Move to gpt-5.6-sol
Fri Oct 23 OpenAI API shutdown: o1, o1-pro, o3-mini, o4-mini, plus legacy GPT snapshots Anyone calling those models by API Repoint to a supported model now; nine days is not much

Action Items

Things your engineering team should look at or do this week.

  • Grep your repos and gateway logs for gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini (including fine-tunedTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ variants), and schedule the move to gpt-6-luna before Jan 14, 2027. Run your eval set on the replacement first; OpenAI's page names it as the successor but says nothing about behavior parity.
  • Audit every npm trusted-publisher binding you own by Friday. Delete and recreate any binding whose repository was renamed, transferred or deleted, and move any publish workflow triggered by issue_comment to push, release or workflow_dispatch.
  • If you use SAP Cloud SDK for Python, upgrade to 0.45.2 or later, skipping 0.53.0 through 0.55.0, and check tenant subdomains against your registered tenants before they reach the SDK. The advisory does not say whether to rotate client credentials; if any untrusted party could set a tenant subdomain, rotate them.
  • Open Docker Hub account settings, revoke personal access tokens you don't recognize, and switch org automation to organization access tokens. Docker emailed exposed accounts on Oct 2; if you got no email, Docker says you were not exposed.
  • Repoint anything still on o1, o3-mini or o4-mini before Oct 23, and anything pinned to gpt-5.5 in Codex today.

Models & Tools

  • Microsoft Maia 300 inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ accelerator (Oct 12) Infra — Third-generation Azure chip on TSMC N2P with eight HBM4 stacks (288 GB, 18 TB/s per accelerator), 96 accelerators in a 167 kW liquid-cooled rack, and up to 128 racks (12,288 accelerators) in a superpod. Microsoft claims up to 7x Maia 200's rack-scale token throughput and 40% more tokens per megawatt than unnamed GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ systems; both are vendor-reported and unbenchmarked. Production deployments start in 2027, with no pricing, regions or GA date; Ignite (Nov 17–20) is the next detail drop. The practical read: this is a cost lever for Microsoft's own inference, not something to plan capacity around this year. (Microsoft | Azure blog | The Register)
  • Google Play lets anyone ship private, AI Studio-built native apps to their own device (Oct 13) Agents — Built in Google AI Studio and installed through Play with a normal Google account, the apps stay invisible in search and usable only by their creator. US-only, 18+ at launch; sharing with trusted friends and family is promised "in the coming months," and going public requires the usual trust-and-safety review. Worth watching as the first mainstream path from vibe-coded prototype to native Android app, but the article gives no developer terms, fees or limits. (TechCrunch)
  • Apple "Minions": a virtual GPU cloud pooled from idle Macs (Oct 13) Infra — Apple researchers describe a server, PostgreSQL metadata store, per-Mac executors and a watchdog that reassigns work when a machine drops. Apple says inference queue times fell from over ten hours to minutes. It is an internal research system with no code release or access program, so treat it as a design to borrow, not a product. (Apple ML Research | The Register)
  • Industrial Next demos reconfigurable robot stations at TechCrunch Disrupt (Oct 13) Robotics/IoT — The San Francisco startup claims its foundation-model-driven stations can be retooled "in minutes rather than weeks" and peel film from camera modules at or above human speed with 100% accuracy (company claims, stage demo). BMW's technology office is evaluating it; the one production customer, a phone maker, is undisclosed. No funding figure was given. (TechCrunch)

APIs & Pricing

  • OpenAI sets Jan 14, 2027 shutdown for the GPT-4o and GPT-4.1 families (Oct 14) Pricing — gpt-4o (2024-08-06 and 2024-11-20 snapshots), gpt-4o-mini, gpt-4.1 and gpt-4.1-mini all leave the API, and fine-tuned versions follow the same date. OpenAI's recommended replacement for all four is gpt-6-luna. This is the broadest OpenAI retirement since the o-series batch; three months is a normal eval-and-migrate window, but fine-tune owners should confirm now whether a tuning path exists on the successor. The page covers API access only and says nothing about ChatGPT. (OpenAI deprecations)
  • Claude Haiku 5.5: $0.10 in / $0.50 out per million tokens (Oct 7; not previously covered) Pricing — Anthropic says that is roughly 75% below Haiku 4.5 and prices cache reads at $0.01; prompts over 100K tokens are $0.50/$2.50. It also halved Sonnet 5.5's cache-read price to $0.10. Every benchmark in the launch (for example Terminal-Bench 4.0 at 39.2% versus 70.6% for Sonnet 5.5) is Anthropic-reported. Anthropic itself points it at summarization, compaction and subagent work rather than hard agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ coding. Available on AWS, Google Cloud and Azure as claude-haiku-5-5. Late catch: a week old, included because it changes subagent cost math. (Anthropic | GitHub changelog)
  • SpaceXSI folds Grok, Cursor and X Premium into one "X Subscription" (Oct 13) Pricing Agents — Five tiers: Free, Plus $8, Premium $30, Super $100, Ultra $200 per month, with AI usage caps that rise by tier (caps not published). The company says existing X Premium+ subscribers pay the same or less. Nothing in the report says what happens to standalone Cursor plans or limits, so Cursor customers should wait for Cursor's own terms before assuming anything. (TechCrunch)

Security

  • SAP Cloud SDK for Python leaks OAuth2 client credentials, CVE-2026-89111 (Oct 13) Security — Critical · CVSS 10.0 · Not exploited (none reported) · Patched in 0.45.2 and later. The SDK builds its token URL from the tenant subdomain without validating it, so an attacker who controls the subdomain can redirect the request, with the app's client ID and secret, to their own server. Multi-tenant apps on SAP BTP are the exposed shape. The Hacker News found in its own testing that releases 0.53.0 through 0.55.0 skip the check for Destination calls; the advisory (GHSA-4mv3-jxxm-9cx2) does not say whether those fall under the CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗. Workaround if you can't upgrade: validate subdomains against registered tenants. (The Hacker News)
  • Docker Hub "GhostPat": converting an account to an organization transferred its access tokens, CVE-2026-105445 (Oct 12–13) Security — Critical · CVSS 9.4 (CVSS 4.0) · Not exploited evidence found by Docker · Fixed server-side on Sep 30, no client update needed. For about two years from Aug 26, 2024, tokens followed the account to whoever was named owner, with no confirmation or notification, giving registry push, pull and delete as the original user. Docker reviewed roughly 304,000 pushes made with transferred tokens and found they all went to the converted org or its named owner; its advisory reports no equivalent result for pulls or deletes. Docker deactivated transferred tokens Oct 8. (The Hacker News)
  • npm trusted publishing hijack "SlugFest": 1,500+ packages exposed (Oct 13; GitHub fix Oct 2) Security — High (our rating) · CVSS: not yet assigned, no CVE, so scanners will be silent · Exploitation unknown (none reported) · Mitigated platform-side Oct 2; maintainers must still audit. Trusted publishing matched a repo by name, not numeric ID, until the first successful publish, so an attacker who re-registered a freed name could mint a valid publish credential for a package they don't own. Bloom Security, which reported it June 6, counts 1,500+ vulnerable packages with 10 million+ combined downloads. GitHub now expires unvalidated bindings after 48 hours and rejects trusted-publishing tokens from issue_comment events. (GitHub changelog | The Hacker News)

Industry & Business

  • Jefferies: cheap Chinese and open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ models could force an AI infrastructure "reset" (Oct 13) Open Source — The analyst note, as relayed by The Register, says roughly half of the ~$2.3 trillion contracted cloud backlog at Microsoft, Alphabet, Amazon and Oracle belongs to OpenAI and Anthropic, and that Moonshot's open-weight Kimi K3 reaches about 95% of Claude Fable 5's performance (Jefferies' estimate). Citing OpenRouter data, it says Chinese models processed 76.1 trillion tokens in the week to Sep 27 versus 14.2 trillion for the top four US providers. These are analyst figures and the note itself is not linked, but the usage split is the number to check against your own routing. (The Register)
  • Pluto Security raises $42M to police employees' AI tools and agents (Oct 13) Security — An $8.5M seed plus a $34M Series A (Jump Capital led the A). It inventories AI tools, skills, extensions and MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ servers on endpoints and can stop an agent mid-task when it reaches for permissions beyond what the user intended. The company reports over $2M ARR and 30+ enterprise customers since selling began in April. It also offers a free scanner for public skills and connectors. (TechCrunch)
  • Judge orders OpenAI to hand Apple hiring records in trade-secrets case (Oct 13) Hiring — Judge Edward Davila granted expedited access to interview materials for ex-Apple staff who applied to OpenAI's hardware and robotics groups, plus private LINE messages between named employees, and approved depositions of three people including chief hardware officer Tang Yew Tan. He denied access to OpenAI's hardware R&D, without prejudice. OpenAI has five days for materials already collected and 14 days for the rest. Apple filed in July. (TechCrunch)

AI in Production

  • Anthropic replaced Git with Scion because agent-scale merges overwhelmed it (Oct 13) Production Agents — Merged changes to its monorepo rose from about 2,500 a day in May to over 30,000 by September; the main history is near 100 GB and CI runs about 25 million checkouts a day. Scion stores server-side changesets in a write-ahead log with compare-and-swap ref updates, and clients keep only a bounded cache. Anthropic reports a roughly 200x smaller CPU and memory footprint than its old mirror fleet, CI and research checkouts dropping from minutes to seconds, and 99%+ cache hit rates (all self-reported). The part to steal: processes built for humans, such as global merge-queue locks, break at agent volume, and Anthropic's answer was post-merge and pre-deploy testing. Scion will not be open-sourced and Anthropic says Git remains right for most codebases. (Anthropic Engineering | The Register)
  • OpenAI's always-on "dot" assistant used ~307K tokens in a day with no prompts (Oct 13) Agents Production — Agent-security firm Geordie logged 155 agent traces in one day on a read-only setup, at least 50 of them inside a two-minute window, and The Register estimates the cost at $5 to $10 per ~200K tokens on OpenAI's Astra model. No OpenAI response was reported and Geordie's data is not published, so this is one vendor's test, but it is the budget question to ask before piloting any background agent: what does it cost when nobody is typing? (The Register)
  • GitHub secret scanning now blocks Anthropic API keys by default (Oct 13) Security — Push protection now covers anthropic_service_api_key and anthropic_user_api_key automatically in repositories with secret scanning on, including free public repos. GitHub also added detection for Cloudsmith keys and made Salesforce Marketing Cloud OAuth tokens a partner pattern. The changelog does not describe an opt-out, so expect surprise blocked pushes in test fixtures and docs that contain example keys. (GitHub changelog)

Policy & Safety

  • Anthropic: small models may be months from self-replicating across cheap IoT devices (Oct 12) Agents Policy — The Frontier Red Team tested open-weight models: GLM-5.2 spread to all 8 nodes in a firewalled emulated network in 4 of 5 trials, using a deliberately trivial exploit (an open ADB port), so Anthropic calls the result an upper bound. A physical demo ran Gemma4-26B across 15 Android TV boxes at 0.39 tokens a second. The "months away" timeline rests on extrapolating a small-model capability trend (R² 0.95 on a rough proxy), and The Register calls the threat speculative. Defender advice is network-level: flow monitoring for multi-gigabyte model transfers, honeypots, quarantine with subscriber notification, and stronger IoT default-credential rules. For an engineering team the actionable piece is smaller: don't leave debug ports like ADB exposed on anything you ship. (Anthropic | The Hacker News | The Register)

Later update (08:31 CT)

Six items that were missing from the early-morning edition. Security and agent-capability stories dominate again; the thread is that AI-assisted bug hunting is now producing volume (a record Patch Tuesday) and depth (a full Chrome chain) at the same time.

Security

  • XBOW built a full Chrome exploit chain with a public OpenAI model for about $40,000 in six days (Oct 13) Security Agents — High (our rating; already fixed) · CVEs: CVE-2026-85045 (V8) and CVE-2026-78952 (Crashpad) per The Hacker News, the WebP decoder bug has no public ID · Not exploited (XBOW says the chain was never used against users) · Fixed in Chrome 152.0.7977.82 and later. Agents running OpenAI's publicly available GPT-5.6 Sol, directed by four XBOW researchers, chained three bugs into code execution on Windows outside Chrome's sandboxes from just loading a page. Google paid $250,000, the top full-chain reward. The cost line is the load-bearing one: XBOW contrasts six days and ~$40K of compute with the roughly six months it says a state-level team once needed. Both figures are XBOW's own. Action: confirm fleet Chrome is on 152.0.7977.82 or later and relaunched. (XBOW | The Hacker News)
  • Linux kernel "Vghost" (CVE-2026-74580): local root and, in one configuration, guest-to-host escape, found by Tencent's Corvus multi-agent system (Oct 14; fix Aug) Security Infra — High · CVSS 8.8 (kernel.org CNA; Red Hat scores it 7.3) · Not exploited public sign · Patched in 5.10.265, 5.15.216, 6.1.183, 6.6.152, 6.12.104, 6.18.45, 7.1.9 and 7.2. The vhost metadata cache isn't reset when a live vring is reconfigured, so stale IOTLB mappings allow out-of-bounds host memory access. Tencent's Zhuque Lab reproduced local-to-root on Ubuntu 26.04, RHEL 9.8, Amazon Linux 2023, Azure Linux 3.0 and Oracle Linux 10.2. The guest-escape path needs a virtio device behind a virtual IOMMU with vhost in device-IOTLB mode, and The Hacker News reports no interim workaround for it. As of Oct 13 Ubuntu 24.04 and 26.04 fixes were still in progress. Until patched, restrict access to /dev/vhost-net and /dev/vhost-vsock. (NVD/CVE record | The Hacker News)
  • Microsoft's October Patch Tuesday is a record 1,411 flaws, and commentators tie it to AI-assisted bug finding (Oct 13) Security Trending — High · Exploited in the wild (CVE-2026-96951, Windows Shell Mark-of-the-Web bypass, CVSS 8.8, on CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗'s KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ with a federal deadline of Oct 16) · Patched in the October cumulative updates. BleepingComputer counts 1,411 (70 Critical, 3 zero-daysAn attack using a flaw that was unknown before it was exploited.NIST ↗); The Hacker News counts 1,413, so treat the exact number as approximate. Two zero-days were publicly disclosed before the patch: CVE-2026-108398 (.NET Framework, Windows Defender Application Control bypass) and CVE-2026-92707 (MessagePack-CSharp denial of service). Zero Day Initiative's Dustin Childs says AI is "partly" why volume is up, a causal claim the articles do not demonstrate. The practical effect is patch-queue triage: expect your Windows and .NET fleets to need a defined fast lane. (BleepingComputer | The Hacker News)

Models & Tools

  • GitHub adds AI Scan metrics to security overview (public preview) (Oct 13) Security Agents — Org and enterprise admins can now see AI Scan findings generated and resolved, remediation rate, which repositories produce findings, and how often developers accept autofix suggestions. The "CodeQL pull requests" page becomes "Code scanning on pull requests" with separate CodeQL and AI Scan tabs. It requires GitHub Code Security or Advanced Security and is not available on GitHub Enterprise Server. If you run AI Scan, the autofix acceptance rate is the number that tells you whether it's saving engineer time or generating noise. (GitHub changelog)

Industry & Business

  • Anthropic opens a Singapore office and signs an AI safety MOU with the government (Oct 14) Customer AI Policy — The fifth Asia-Pacific office comes with an MOU with Singapore's Ministry of Digital Development and Information covering safety, security and model evaluation, including technical exchanges on testing standards and incident response. Anthropic's own adoption figures: more than 1,500 GovTech Singapore engineers and technical staff use Claude Code, and Grab's Claude-based sales assistant saves about 40,000 hours a quarter (both Anthropic-reported). For teams in regulated Asia-Pacific markets, the useful detail is in-region Claude inference on Amazon Bedrock. (Anthropic)
  • FBI arrests a former OpenAI product employee for allegedly planting a hidden recording device in an executive conference room (Oct 13) Trending Hiring — Prosecutors charge Abdulrahman Alfozan with intercepting oral communications (up to five years). OpenAI says it found the device about a week before his Oct 9 arrest, used its internal AI models to analyze it, and has no indication that model weights, source code or customer data were accessed. Motive is unclear; FBI Director Kash Patel's statement ties it to the company's "super intelligence technology." Included at [Trending] because the field is discussing it; the engineering takeaway is insider physical-access risk at AI labs, not a vulnerability to patch. (TechCrunch | BleepingComputer)

Type to search · ↑↓ to move · enter to open