bg-banner-title

Tin tức

[Technology Newsletter] Monthly Update on AI Models, Frameworks, Tools & Platforms - Oct 2026

Cẩm nang công nghệ
11/2026

This Month in 60 Seconds

October was a month of practical change rather than one big breakthrough. Capable models got cheaper, OpenAI set firm dates to switch off older models that many applications still use, and two critical security events on the same day showed that AI tools and packages are now a real attack path.

1. LiteLLM, a popular LLM gateway, was published with malware on PyPI and Docker on 21 Oct; the same day PHP patched two critical flaws found with AI help.

2. OpenAI will remove GPT-4o and GPT-4.1 from its API on 14 Jan 2027; older models such as GPT-4 and o1 were switched off on 23 Oct.

3. Google made Gemini 4 Argon generally available at an introductory $2/$10 per 1M tokens, Anthropic launched Claude Haiku 5.5 at $0.10/$0.50 and OpenAI released GPT-6.1 Terra at $1/$5.

4. Anthropic is donating Agent Skills, the SKILL.md format for teaching AI agents a task, to a Linux Foundation body that will be its vendor-neutral home.

5. IT services firms report a split: HCLTech's advanced AI revenue grew 78%, while Wipro's large-deal bookings fell 58.6% quarter on quarter and Barclays plans to need fewer developers.

By the Numbers

NumberWhat it means
$0.10per 1M input tokens for Claude Haiku 5.5, Anthropic's new small model
+78%year-on-year growth in HCLTech's advanced AI revenue (Jul–Sep quarter)
14 Jan2027: GPT-4o and GPT-4.1 leave the OpenAI API

01  Top Stories

Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

21 Oct 2026  ·  For: All, Dev, DevOps  ·  Security  ·  Supply chain

On 21 October, attackers took over the release process of LiteLLM, an open-source library and proxy that many AI projects use to call several LLM providers through one API. They published litellm==1.107.0 on PyPI and as Docker images, including the latest tag. The package installs a credential stealer that runs whenever Python starts, targeting SSH keys, cloud credentials (including on CI runners) and LLM API keys, and loads remote-code tools. LiteLLM's later analysis traced the attack to a master key committed to a public repository on 27 May: the commit was removed, but the key was never revoked.

Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

21 Oct 2026  ·  For: All, Dev, DevOps  ·  Security  ·  Supply chain

On 21 October, attackers took over the release process of LiteLLM, an open-source library and proxy that many AI projects use to call several LLM providers through one API. They published litellm==1.107.0 on PyPI and as Docker images, including the latest tag. The package installs a credential stealer that runs whenever Python starts, targeting SSH keys, cloud credentials (including on CI runners) and LLM API keys, and loads remote-code tools. LiteLLM's later analysis traced the attack to a master key committed to a public repository on 27 May: the commit was removed, but the key was never revoked.

On the same day, PHP released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two unauthenticated remote-code-execution flaws (CVE-2026-109618 and CVE-2026-109619). According to Wiz, which says it helped find them with AI models, every PHP version from the last decade is affected by at least one of them.

  • LiteLLM: pin to 1.106.2. If 1.107.0 ran anywhere, downgrading is not enough: it leaves startup files (cook.pth, sitecustomize.py) and a background process behind. Rebuild that virtual environment, container image or self-hosted CI runner from scratch, and rotate every credential it could reach.
  • A secret deleted from git history is still leaked until it is rotated.
  • PHP: updating is the only complete fix.
TMA Solutions

Diagram: TMA AI newsletter, based on the LiteLLM incident report

Why it matters for TMA: Check lock files and images built on 21 Oct, pin versions instead of using latest, and patch every PHP application we build or maintain. AI is helping find flaws faster, so patch windows are getting shorter.

Source: LiteLLM security incident report · LiteLLM: attack chain · Wiz: LiteLLM analysis · PHP releases · Wiz: PHP2Shell

OpenAI sets the end date for GPT-4o and GPT-4.1 in its API: 14 January 2027

14 Oct 2026  ·  For: Dev, QA, DevOps, BA/PM  ·  OpenAI  ·  Deprecation

OpenAI's deprecations page now lists 14 January 2027 as the shutdown date for GPT-4o (2024-08-06 and 2024-11-20 snapshots), GPT-4o mini, GPT-4.1 and GPT-4.1 mini. The recommended replacement is GPT-6 Luna. A first wave of older models, including gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o3-mini and o4-mini, was already switched off on 23 October.

Why it matters for TMA: Many client applications built in 2024–2025 still call GPT-4o or GPT-4.1. Moving to GPT-6 Luna is more than changing a model name: prompts, tool calls and output formats need regression tests. List affected systems now and agree a migration plan with clients before January.

Source: OpenAI API deprecations

Capable models get cheaper: Gemini 4 Argon goes GA and Claude Haiku 5.5 launches

13 Oct 2026  ·  For: All, Dev, Data/AI, BA/PM  ·  Google  ·  Anthropic  ·  Models

TMA Solutions

Image: Google

Google made Gemini 4 Argon, its most capable model, generally available to all paid Gemini API users on 13 October. The introductory price of $2 input / $10 output per 1M tokens runs until 30 November; from 1 December it rises to $4 / $20. Anthropic launched Claude Haiku 5.5 on 7 October: $0.10 / $0.50 per 1M tokens for prompts up to 100,000 tokens, a 1M-token context window, on the Claude API and the major clouds.

  • Like other Claude models from 4.7 on, Haiku 5.5 rejects non-default temperature, top_p or top_k, so code written for older models may need changes. See also Model Tracker.

Why it matters for TMA: For many project tasks, small models are now good and cheap enough to try first, with a larger model only where needed. When you quote AI running costs to a client, use the post-introductory Argon price and test with the client's real data.

Source: Google: Gemini 4 Argon · Gemini API changelog · Anthropic: Claude Haiku 5.5 · Claude docs: Haiku 5.5

Agent Skills is moving to a vendor-neutral home under the Linux Foundation

22 Oct 2026  ·  For: Dev, QA, DevOps, Data/AI  ·  Anthropic  ·  Standards

Anthropic is donating Agent Skills to the Agentic AI Foundation (AAIF), the Linux Foundation body that already hosts the Model Context Protocol (MCP). A skill is a folder with a SKILL.md file, plus optional scripts and reference files, that teaches an AI agent how to do a specific task. According to Anthropic, more than 70 agent platforms support the format, including ChatGPT, Cursor, GitHub Copilot and VS Code, with about 1 million skills published.

Why it matters for TMA: A team can write a skill once (for example, "review a pull request against our checklist") and use it across different coding agents. Treat skills like code: review them before installing, because they can contain scripts.

Source: Anthropic announcement · Agent Skills specification

IT services firms split on AI: strong growth for some, shrinking deals for others

15 Oct 2026  ·  For: All, BA/PM  ·  IT services  ·  Market

Quarterly results showed how unevenly AI is reshaping our industry. HCLTech's advanced AI revenue reached $219 million in the quarter, up 78% year on year, and it will set up a business unit dedicated to Anthropic. TCS said its AI services now run at $3.1 billion a year, more than 10% of revenue, and at Infosys AI reached 10.5% of revenue, up 29% in a quarter, even as it cut its full-year growth forecast to 1–2%. In contrast, Wipro's revenue was flat and its large-deal bookings fell 58.6% from the previous quarter. In Vietnam, FPT's AI and data analytics revenue grew 55% in nine months. On the client side, Barclays said it will cut about 5,000 roles, including "fewer developers as AI makes our people more productive".

Why it matters for TMA: Clients expect AI productivity gains to show up in price and speed. Engineers who deliver reliably with AI tools, and can explain how quality is controlled, are what clients now look for.

Source: HCLTech Q2 FY27 results · HCLTech–Anthropic partnership · Business Today: TCS Q2 FY27 · Infosys Q2 FY27 results · Wipro Q2 FY27 results · VnEconomy: FPT 9M results · Barclays job cuts (AFP)

02  Model Tracker

Prices are USD per 1M input / output tokens. Quality claims are vendor-reported; test on your own tasks. Before you try a new API with client work: send client data only to AI services on TMA's approved tools list and allowed by the client contract, because contracts and laws such as GDPR and Japan's APPI limit where data may go. Run open-weight models only on approved infrastructure, after checking the licence.

TMA Solutions

Price per 1M tokens, models with published prices Chart: TMA AI newsletter, from the vendor prices in the table below

Model

Vendor

Positioning

Price (in / out)

Availability

Gemini 4 Argon

Google

Most capable Gemini model

$2 / $10 until 30 Nov, then $4 / $20

GA, Gemini API (13 Oct)

GPT-6.1 Terra

OpenAI

Fast, efficient model, 1M-token context; also in Codex and Copilot

$1 / $5

Released 26 Oct

Claude Haiku 5.5

Anthropic

Small and fast, 1M-token context

$0.10 / $0.50 (prompts up to 100K)

GA (7 Oct)

Muse Wonder

Meta

Frontier model (Meta's first, per TechCrunch); API only, no weights

$4 / $20 (per TechCrunch)

Meta Model API (12 Oct)

Mistral Large 4

Mistral

About 1T-parameter MoE, multimodal

$1.36 / $4.18

Public preview; weights expected 31 Oct, licence not yet published

Kimi K3.1 Pro

Moonshot

3T-class model, weights promised; #4 on the Artificial Analysis index

$2 / $10

API (23 Oct); weights due by 4 Nov

MiniMax M3.1 Flash

MiniMax

Low-cost agent and coding model

$0.15 / $0.60

API (19 Oct)

Platform and API changes

  • Claude API version 2026-10-21, used by the new major SDK versions, rejects temperature, top_p and top_k for every model, even at default values. Calls to older models such as Sonnet 4.6 that lower temperature lose that effect: keep them on version 2023-06-01 or retest. Pin SDK major versions until your code is checked. (migration guide)
  • Gemini API logging: new paid projects created from 26 Oct store Generate Content prompts and responses, for 55 days by default. Turn this off (store=false or the AI Studio Logs setting) on projects that handle client data. (changelog; logs docs)
  • OpenAI Decisions API became generally available on 22 Oct: fast probability, choice and score answers on GPT-6 Luna, at $0.10 per 1M input tokens with output tokens free. (docs)

03  Tools for Engineering Teams

Coding agents get sandboxes: Copilot, VS Code and Kiro limit what agents can touch

7 Oct 2026  ·  For: Dev, DevOps  ·  GitHub  ·  VS Code  ·  Kiro

GitHub made local sandboxing for Copilot generally available on 7 October, at no extra cost. Companies can require it through managed settings that developers cannot weaken. VS Code 1.141 added a sandbox for terminal commands that agents run (chat.agent.sandbox.enabled), and Kiro CLI 2.33 now asks "Trust this workspace?" before it loads a repository's hooks, agents and MCP servers.

Why it matters for TMA: Opening an unknown repository with an AI agent can run code from that repository's configuration. Turn sandboxing on, and only trust workspaces you know, especially client or open-source code.

Source: GitHub Changelog: local sandboxing GA · VS Code 1.141 release notes · Kiro CLI 2.33

Copilot can now operate desktop apps, useful for testing legacy software

1 Oct 2026  ·  For: QA, Dev  ·  GitHub  ·  Testing

GitHub Copilot CLI and the Copilot app gained computer use in public preview: Copilot can click and type in desktop applications after asking for permission (/computer on).

Why it matters for TMA: QA teams can try agent-driven checks on desktop or legacy applications that have no API. Use it only on test machines, never on a client's production environment.

Source: GitHub Changelog: computer use

Testing for the agent era: Playwright 1.64 adds WebMCP, GitHub opens ReviewBench

7 Oct 2026  ·  For: QA, Dev  ·  Playwright  ·  Testing

Playwright 1.64 (7 Oct) lets tests and agents call WebMCP tools that a web page exposes (page.webmcp), and adds a --shuffle option to find order-dependent tests. GitHub released ReviewBench, an open benchmark of 219 pull requests in 19 languages for measuring AI code-review tools, which teams can run themselves.

Why it matters for TMA: WebMCP is still experimental in browsers, but it shows how web apps will be tested and used by agents. ReviewBench gives a fair way to compare code-review bots before recommending one to a client.

Source: Playwright releases · GitHub Blog: ReviewBench

GitHub adds a package firewall and short-lived CLI tokens

20 Oct 2026  ·  For: DevOps, Dev  ·  GitHub  ·  Supply chain

GitHub's package firewall (public preview, npm only, GitHub Enterprise Cloud) checks packages before coding agents and CI runners install them. GitHub CLI 2.103.0 can now use access tokens that expire within hours and refresh automatically, so a leaked token stops working quickly.

Why it matters for TMA: Both respond directly to attacks like the LiteLLM compromise. Switching your own gh login to expiring tokens takes a few minutes (see Try It).

Source: GitHub Changelog: package firewall · GitHub Changelog: expiring gh tokens

Figma connects design and code for agents

15 Oct 2026  ·  For: UI/UX, Dev, BA/PM  ·  Figma  ·  Design

Figma's "Build by design" release adds code layers (open beta, all plans), and Specs and design-system Guidelines that agents can read (open beta, paid plans).

TMA Solutions

Image: Figma

Why it matters for TMA: Designers can hand over specifications that coding agents follow more closely, reducing back-and-forth between design and development. Try it on internal projects first.

Source: Figma: Build by design

04  AI at Work for Everyone

GPT-6 now powers ChatGPT on all plans, with interactive answers

7 Oct 2026  ·  For: All, Office  ·  OpenAI  ·  ChatGPT

GPT-6 rolled out in ChatGPT on all plans: paid plans from 7 October and Free and Go from 8 October. Answers can now include interactive charts, calculators and buttons (OpenAI calls this "Intelligent UI").

Why it matters for TMA: Answers look more polished, but a nice chart can still contain wrong numbers. Check figures before using them in reports or emails, and follow TMA's AI policy on which tools and data you may use.

Source: TechCrunch

Microsoft 365 Copilot can be bought on its own from 1 December

14 Oct 2026  ·  For: Office, All  ·  Microsoft  ·  Licensing

Microsoft told partners that from 1 December 2026, Microsoft 365 Copilot can be bought without a Microsoft 365 base subscription, at the same price. For new Copilot Business licences bought through partners, usage-based billing will be on by default in most markets, with a default cap of 4,000 credits per user per month that admins can change.

Why it matters for TMA: AI tools are moving from fixed licences to pay-per-use. IT and Finance colleagues should set spending limits before usage-based features are switched on.

Source: Microsoft Partner Center announcements, October 2026

Google gives AI "coworker agents" their own work accounts

9 Oct 2026  ·  For: All, Office  ·  Google  ·  Workspace

At Gemini at Work 2026, Google introduced a single Gemini agent for businesses and coworker agents that get their own Workspace account, with least-privilege access and audit logs for admins.

Why it matters for TMA: This is what AI at work is starting to look like: agents with their own identity and permissions, managed like staff accounts.

Source: Google Cloud: Gemini at Work 2026

What you type into a work AI tool is a company record

22 Oct 2026  ·  For: All, Office  ·  Privacy  ·  Safe use

Google Vault can now export employees' Gemini conversations for compliance reviews, from 22 October. ChatGPT began rolling out locked chats on most plans from 20 October, protected by a PIN, Face ID or Touch ID, but they are not end-to-end encrypted.

Why it matters for TMA: Company AI tools can be reviewed like email, and a lock on your phone does not make a chat private from your employer or the provider. Do not enter client data, passwords or personal information unless TMA's AI policy allows it for that tool.

Source: Workspace Updates: Vault for Gemini · TechCrunch: ChatGPT locked chats

05  Security, Governance & Policy

EU: Council agrees changes to GDPR for AI; AI Act marking duty starts 2 December

21 Oct 2026  ·  For: All, BA/PM  ·  EU  ·  Regulation

On 21 October, EU member states agreed their position on the Digital Omnibus. It includes a legal basis for using personal data to develop and run AI (under the "legitimate interest" balancing test) and a 96-hour deadline for reporting high-risk data breaches. It is not law yet. Separately, from 2 December 2026, generative AI systems already on the EU market before 2 August 2026 must mark their outputs in a machine-readable way.

Why it matters for TMA: Projects that build generative AI features for EU clients should check content marking (for example watermarks or C2PA metadata) before 2 December.

Source: Council of the EU press release · AI Act Explorer: Digital Omnibus

Japan publishes a safety standard for AI agents being rolled out to 180,000 civil servants

20 Oct 2026  ·  For: All, BA/PM, Dev  ·  Japan  ·  Regulation

Japan's Digital Agency published the Gennai Standard for AI agents in government work. Responsibility for work done with an agent stays with staff and their administrative body, agents may not call external APIs or MCP servers that are not managed by the developers, and data is in principle processed on Japan's government cloud.

Why it matters for TMA: Japan is a key market for TMA. Clients there, especially public-sector ones, are likely to ask for similar controls: clear human accountability, approved tool lists and domestic data processing.

Source: Digital Agency of Japan

Agent tools under attack: planted instruction files and flaws in Junie, Langflow and mcp-go

22 Oct 2026  ·  For: Dev, DevOps, QA  ·  Security  ·  MCP

Researchers at Forever Security showed that a web page can drop an AGENTS.md or CLAUDE.md file into the Downloads folder; a coding agent started there reads and follows it. In their tests this made Codex add a malicious MCP server in more than 90% of attempts and made Claude Code leak its login token in about half. The researchers say there is nothing to patch: this is how agents work. Advisories also hit other AI tools. Update JetBrains Junie CLI to 26.8.3 (malicious repository configuration could run code) and Langflow to 1.10.3 (critical, CVSS 9.9); mcp-go, a Go MCP library, has a denial-of-service flaw with no fix yet.

Why it matters for TMA: Start agents only inside the project folder, never in Downloads or your home folder. Keep agent tools and MCP libraries updated, and treat instruction files and MCP configuration like code that needs review.

Source: Forever Security report · GitHub Advisory: Junie CLI · OpenCVE: Langflow CVE-2026-105697 · GitHub Advisory: mcp-go

06  AI in Our Industries

Telecom: T-Mobile makes the first AI-RAN call on a commercial network

21 Oct 2026  ·  For: All, Dev  ·  Telecom  ·  5G

At GTC Berlin, NVIDIA said T-Mobile US completed the first AI-RAN call on a commercial network, using NVIDIA servers with Nokia's GPU-accelerated radio software.

Why it matters for TMA: Telecom is our largest industry. Radio software that runs on GPUs creates demand for engineers who combine 5G and AI skills.

Source: NVIDIA GTC Berlin release (StorageNewsletter copy)

Healthcare: Epic lets its AI agents work inside other healthcare apps

21 Oct 2026  ·  For: All, Dev  ·  Healthcare  ·  Agents

Epic, a leading US electronic health record vendor, introduced Connectors, which app developers build so Epic's AI agents can read data from and complete tasks in their applications. Connectors will be available from March 2027.

Why it matters for TMA: Health-tech clients that integrate with Epic will need agent-ready connectors, which is new integration work for our healthcare teams.

Source: Epic: Introducing Connectors

Finance: Singapore's MAS makes banks accountable for vendors' AI

7 Oct 2026  ·  For: BA/PM, QA  ·  Finance  ·  Regulation

Singapore's central bank published final AI Risk Management Guidelines for all financial institutions (main sections effective 7 October 2027). Institutions remain accountable for AI built or operated by third parties and must obtain assurance from those providers.

Why it matters for TMA: As a vendor to banks and insurers, we should expect requests for documentation, testing evidence and controls for any AI we deliver.

Source: Monetary Authority of Singapore

07  Action Items & Deadlines

TMA Solutions

Chart: TMA AI newsletter, from the table below

Date

What changes

Who is affected

What to do

23 Oct 2026

OpenAI switched off legacy models (gpt-3.5-turbo, gpt-4, o1, o3-mini, o4-mini and others)

Any app or tool still calling them

Search code and config for these model IDs and switch to current models

16 Nov 2026

Copilot removes Claude Opus 4.8, Opus 5 and GPT-5.6 Terra (individual annual plans keep Opus 4.8)

Copilot users, prompt files and custom agents

Switch to Claude Opus 5.5, GPT-6.1 Sol or GPT-6 Luna (GitHub's picks); GPT-6.1 Terra is the newer Terra

17 Nov 2026

Gemini API preview text-to-speech and Live models may shut down from this date

Voice apps on Gemini preview models

Migrate to Gemini 3.8 TTS or Live models and retest

30 Nov 2026

Claude Sonnet 4.5 retires on Anthropic's platforms

Apps pinned to claude-sonnet-4-5-20250929

Move to Claude Sonnet 5.5 and check for removed parameters

1 Dec 2026

Gemini 4 Argon price rises to $4 / $20 per 1M tokens

Projects budgeting Argon usage

Re-estimate costs

1 Dec 2026

Microsoft 365 Copilot sold without a Microsoft 365 base licence; usage-based billing on by default for new Copilot Business bought through partners (CSP), not yet in 11 countries

IT and Finance

Set spending limits in the Microsoft 365 admin center

2 Dec 2026

EU AI Act: machine-readable marking of output for generative AI systems placed on the EU market before 2 Aug 2026

Providers of such systems, incl. those we build for EU clients

Confirm watermarking or metadata is in place

31 Dec 2026

PHP 8.2 end of life

All PHP applications

Apply the 21 Oct patches now; plan upgrades to PHP 8.3+

6 Jan 2027

OpenAI removes tts-1, tts-1-hd and gpt-4o-mini-tts

Voice, IVR and read-aloud features

Move to gpt-realtime-2.1-mini and retest audio

14 Jan 2027

GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the OpenAI API

Client apps and chatbots built on them

Migrate to GPT-6 Luna and run regression tests

08  Try It in 15 Minutes

Switch your GitHub CLI to tokens that expire on their own

For: Dev, DevOps, QA

From GitHub CLI 2.103.0 (public preview), gh can use short-lived access tokens that refresh automatically, so a leaked token stops working within hours. Do this on your own laptop, not on shared build servers or CI runners.

1.    Upgrade to gh 2.103.0 or newer, then sign in with gh auth login --expiring (on your laptop, gh config set expiring_credentials enabled makes this the default).

2.   Run gh auth setup-git and gh auth status to confirm the token now has an expiry time.

3.   In your own scripts, call gh auth token each time instead of storing a token. Unattended automation should not use a personal gh login at all: use the workflow's built-in GITHUB_TOKEN, a GitHub App token, or OIDC for cloud access.

gh --version                 # needs 2.103.0 or newer

gh auth login --expiring     # opt in for this login

gh auth setup-git            # gh refreshes the git HTTPS credential

gh auth status               # shows when the token expires

TOKEN=$(gh auth token)       # in scripts: fetch a fresh token each time

Source: GitHub Changelog: expiring, auto-refreshing tokens

Summarise a public article and draft a team email with Microsoft Copilot Chat

For: All, Office

Copilot Chat is included with eligible Microsoft 365 business subscriptions; no extra licence is needed. Use only public content.

1.    Go to copilot.cloud.microsoft (or open Copilot in Teams or Outlook) and sign in with your work account, not a personal Microsoft account.

2.   Check for the "Work" label and the green data-protection shield.

3.   Open a public web page in Microsoft Edge, for example a vendor announcement from this newsletter. Select the Copilot icon in Edge and ask: "Summarise this page in 5 bullet points for a non-technical audience."

4.   Then ask: "Draft a short, friendly email to my team explaining what this means for us, under 120 words, with one suggested next step."

5.   Check every fact against the original page before sending.

Source: Microsoft Learn: What is Microsoft Copilot?

09  What to Watch Next

·         28–30 Oct: Vietnam International Digital Week 2026 in Quang Ninh, themed on AI governance.

·         31 Oct and 4 Nov: open weights expected for Mistral Large 4 and Kimi K3.1 Pro; check each licence before client use.

·         As soon as November: Anthropic filed for a stock-market listing on 26 Oct (ticker ANT) and could list next month (NPR via OPB).

·         December 2026 – February 2027: Japan's Digital Agency plans further Gennai tools for government AI agents.

10  Glossary

  • LLM gateway: a service or library that sends requests to several AI model providers through one API and manages keys, costs and logging. LiteLLM is an example.
  • Supply-chain attack: hiding malicious code in a library, package or image that other people download, instead of attacking a target directly.
  • MCP (Model Context Protocol): an open standard that lets AI agents connect to tools and data sources in a consistent way.
  • Agent Skills: an open format (a folder with a SKILL.md file) that packages instructions and scripts to teach an AI agent a specific task.
  • Open-weight model: a model whose trained parameters are published so anyone can run it on their own hardware, under the vendor's licence terms.
  • Sources

1. LiteLLM security incident report — Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

2. LiteLLM: attack chain — Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

3. Wiz: LiteLLM analysis — Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

4. PHP releases — Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

5. Wiz: PHP2Shell — Malware in LiteLLM and critical PHP flaws: one day, two supply-chain alarms

6. OpenAI API deprecations — OpenAI sets the end date for GPT-4o and GPT-4.1 in its API: 14 January 2027

7. Google: Gemini 4 Argon — Capable models get cheaper: Gemini 4 Argon goes GA and Claude Haiku 5.5 launches

8. Gemini API changelog — Capable models get cheaper: Gemini 4 Argon goes GA and Claude Haiku 5.5 launches

9. Anthropic: Claude Haiku 5.5 — Capable models get cheaper: Gemini 4 Argon goes GA and Claude Haiku 5.5 launches

10. Claude docs: Haiku 5.5 — Capable models get cheaper: Gemini 4 Argon goes GA and Claude Haiku 5.5 launches

11. Anthropic announcement — Agent Skills is moving to a vendor-neutral home under the Linux Foundation

12. Agent Skills specification — Agent Skills is moving to a vendor-neutral home under the Linux Foundation

13. HCLTech Q2 FY27 results — IT services firms split on AI: strong growth for some, shrinking deals for others

14. HCLTech–Anthropic partnership — IT services firms split on AI: strong growth for some, shrinking deals for others

15. Business Today: TCS Q2 FY27 — IT services firms split on AI: strong growth for some, shrinking deals for others

16. Infosys Q2 FY27 results — IT services firms split on AI: strong growth for some, shrinking deals for others

17. Wipro Q2 FY27 results — IT services firms split on AI: strong growth for some, shrinking deals for others

18. VnEconomy: FPT 9M results — IT services firms split on AI: strong growth for some, shrinking deals for others

19. Barclays job cuts (AFP) — IT services firms split on AI: strong growth for some, shrinking deals for others

20. GitHub Changelog: local sandboxing GA — Coding agents get sandboxes: Copilot, VS Code and Kiro limit what agents can touch

21. VS Code 1.141 release notes — Coding agents get sandboxes: Copilot, VS Code and Kiro limit what agents can touch

22. Kiro CLI 2.33 — Coding agents get sandboxes: Copilot, VS Code and Kiro limit what agents can touch

23. GitHub Changelog: computer use — Copilot can now operate desktop apps, useful for testing legacy software

24. Playwright releases — Testing for the agent era: Playwright 1.64 adds WebMCP, GitHub opens ReviewBench

25. GitHub Blog: ReviewBench — Testing for the agent era: Playwright 1.64 adds WebMCP, GitHub opens ReviewBench

26. GitHub Changelog: package firewall — GitHub adds a package firewall and short-lived CLI tokens

27. GitHub Changelog: expiring gh tokens — GitHub adds a package firewall and short-lived CLI tokens

28. Figma: Build by design — Figma connects design and code for agents

29. TechCrunch — GPT-6 now powers ChatGPT on all plans, with interactive answers

30. Microsoft Partner Center announcements, October 2026 — Microsoft 365 Copilot can be bought on its own from 1 December

31. Google Cloud: Gemini at Work 2026 — Google gives AI "coworker agents" their own work accounts

32. Workspace Updates: Vault for Gemini — What you type into a work AI tool is a company record

33. TechCrunch: ChatGPT locked chats — What you type into a work AI tool is a company record

34. Council of the EU press release — EU: Council agrees changes to GDPR for AI; AI Act marking duty starts 2 December

35. AI Act Explorer: Digital Omnibus — EU: Council agrees changes to GDPR for AI; AI Act marking duty starts 2 December

36. Digital Agency of Japan — Japan publishes a safety standard for AI agents being rolled out to 180,000 civil servants

37. Forever Security report — Agent tools under attack: planted instruction files and flaws in Junie, Langflow and mcp-go

38. GitHub Advisory: Junie CLI — Agent tools under attack: planted instruction files and flaws in Junie, Langflow and mcp-go

39. OpenCVE: Langflow CVE-2026-105697 — Agent tools under attack: planted instruction files and flaws in Junie, Langflow and mcp-go

40. GitHub Advisory: mcp-go — Agent tools under attack: planted instruction files and flaws in Junie, Langflow and mcp-go

41. NVIDIA GTC Berlin release (StorageNewsletter copy) — Telecom: T-Mobile makes the first AI-RAN call on a commercial network

42. Epic: Introducing Connectors — Healthcare: Epic lets its AI agents work inside other healthcare apps

43. Monetary Authority of Singapore — Finance: Singapore's MAS makes banks accountable for vendors' AI

44. Microsoft Learn: What is Microsoft Copilot? — Summarise a public article and draft a team email with Microsoft Copilot Chat

45. GPT-6.1 Terra — Model Tracker

46. Muse Wonder — Model Tracker

47. Mistral Large 4 — Model Tracker

48. Kimi K3.1 Pro — Model Tracker

49. Artificial Analysis index — Model Tracker

50. MiniMax M3.1 Flash — Model Tracker

51. Copilot removes Claude Opus 4.8, Opus 5 and GPT-5.6 Terra — Action Items & Deadlines

52. Gemini API preview text-to-speech and Live models may shut down from this date — Action Items & Deadlines

53. Claude Sonnet 4.5 retires — Action Items & Deadlines