What is happening in the AI world right now — curated, verified, with sources.
Security2026-07-30
Anthropic: Claude penetrated real systems of three companies during security tests
After reviewing 141,006 cybersecurity evaluation runs, Anthropic disclosed three incidents in which Claude models reached the open internet from the test environment of partner Irregular and gained unauthorized access to the production systems of three unrelated organizations. It was not a breakout: the prompt told the models they were in a simulation with no internet access — but due to a misunderstanding between Anthropic and the partner, access was in fact open. The models treated the real targets as part of a Capture-the-Flag exercise and used weak passwords and unprotected endpoints. Opus 4.7, Mythos 5 and an internal test model were affected; Anthropic subsequently halted all cyber evaluations.
Anthropic: Investigating three real-world incidents in our cybersecurity evaluations ↗ · SecurityAffairs: Anthropic Finds Claude Breached Real Companies During Security Evaluations ↗
Tools2026-07-30
VS Code bundles July updates for Copilot: Worktrees, multi-chat, Vision
GitHub summarizes the Copilot updates in VS Code versions 1.127 through 1.131, released in July. The Agents panel gets a redesigned layout with a more compact diff view and now lets Copilot, Claude and Codex sessions optionally run in isolated Git worktrees. New additions include multi-chat sessions with branchable peer chats, image support in chat now generally available, and support for bring-your-own-key (BYOK) models in the Agents panel. Business and Enterprise users can also now see their AI credit usage for the current billing period directly in the Copilot status menu.
GitHub Changelog: GitHub Copilot in Visual Studio Code, July 2026 releases ↗ · Visual Studio Code 1.131 Release Notes ↗
MCP2026-07-29
Copilot code review now integrates Agent Skills and MCP servers as a standard feature
GitHub has ended the preview announced in June: Copilot Code Review now supports both Agent Skills and MCP servers as a standard feature for all Pro, Pro+, Business and Enterprise users. A SKILL.md file under .github/skills lets teams bring in internal tools and standards, while MCP connections pull context from external platforms such as issue trackers into the review; tool calls remain limited to read access. MCP configurations set up for the cloud agent automatically apply as well, and GitHub MCP and Playwright MCP are active by default. Comments will now indicate when they are based on Skills or MCP context.
GitHub Changelog: Copilot code review: Agent skills and MCP now generally available ↗
MCP2026-07-28
MCP specification 2026-07-28 final: protocol becomes stateless
The MCP specification 2026-07-28 has been finalized and breaks with the previous protocol: the session handshake (initialize/initialized) and the Mcp-Session-Id header are removed, MCP becomes stateless and self-describing per request, and servers can now be scaled behind simple round-robin load balancers. Server-side requests such as Sampling, Elicitation and Roots now run via the new Multi-Round-Trip pattern instead of open bidirectional streams. Tasks moves out of the core into its own extension. Roots, Sampling, Logging and the HTTP+SSE transport are now deprecated, with a transition period of at least twelve months. Existing MCP servers and clients must be adapted for the new version.
GitHub Release 2026-07-28 (modelcontextprotocol/modelcontextprotocol) ↗ · MCP Changelog: Key Changes seit 2025-11-25 ↗ · MCP Blog: The 2026-07-28 Specification ↗
Tools2026-07-28
Grok 4.5 arrives in GitHub Copilot
xAI has made its Grok 4.5 model available in GitHub Copilot, selectable in the model picker of VS Code, Visual Studio, Copilot CLI and other Copilot surfaces. Grok 4.5 offers a context window of up to 500,000 tokens, processes text and image input, and can be set to three reasoning levels. GitHub cites internal tests showing good results on terminal-based coding tasks and parallel tool use. The rollout is happening gradually for Copilot Pro, Pro+, Max, Business and Enterprise, billed at xAI's list prices.
GitHub Changelog: Grok 4.5 is now available in GitHub Copilot ↗ · xAI News: Grok 4.5 in GitHub Copilot ↗
Security2026-07-28
npm to automatically scan packages at publish time
npm is introducing automatic malware scanning of newly published packages before they become installable - depending on the finding, packages are released, held back for manual review, or blocked. This usually delays availability by a few minutes, longer during peak times. Suspended maintainers can file an appeal, though additional account measures may follow in some cases. npm is also introducing a contentPolicy field in package.json for legitimate packages that resemble malware, such as security tools. Such dual-use packages require a DISCLOSURE file explaining their purpose and function, which is reviewed by the Trust & Safety team.
GitHub Changelog: npm publish-time malware scanning and dual-use metadata ↗
Security2026-07-27
NVIDIA and Industry Leaders Launch the Open Secure AI Alliance
NVIDIA has founded the Open Secure AI Alliance together with other organizations; the announcement names Microsoft, IBM, Cisco, Cloudflare, CrowdStrike, Databricks, Hugging Face, LangChain, Cognition, and the Linux Foundation among its members. The goal is open collaboration on tools for securing AI agents — identity, access management, isolation, guardrails, logging, and secure coding workflows. The first technical contribution is NOOA (NVIDIA Labs Object-Oriented Agent), an open-source framework under the Apache-2.0 license that models agent behavior as testable, versionable Python code instead of scattered prompts and tool schemas.
Industry Leaders Unite in Open Secure AI Alliance (NVIDIA Blog) ↗ · NVIDIA-NeMo/labs-OO-Agents (GitHub) ↗
Tools2026-07-27
GitHub Copilot App Gets Its Own Access Policy and Central Policy Control
GitHub has added the Copilot app and the Copilot cloud agent to its centrally managed enterprise settings. Organizations can now use a shared managed-settings.json file to define which plugins and marketplaces developers may use and whether approval prompts can be bypassed — previously this applied only to Copilot CLI and VS Code. In addition, the Copilot app now has its own access policy, independent of the CLI policy, at the enterprise and organization level, enabled by default for all users. Admins can allow or block access per organization, or delegate the decision to organization admins.
GitHub Changelog: Enterprise managed settings in the GitHub Copilot app and Copilot cloud agent ↗ · GitHub Changelog: Manage GitHub Copilot app access with a dedicated policy ↗
Models2026-07-27
Kimi K3: Weights Now Actually Downloadable on Hugging Face
Kimi K3 was announced on July 16, but the weights remained unavailable until now. Since July 27, they have been available for download under moonshotai/Kimi-K3 — with no access restrictions, under the model's own Kimi-K3 license. The model card lists 2.8 trillion parameters, Kimi Delta Attention with Attention Residuals, a context window of 1 million tokens, and native processing of text, image, and video within the same model. This makes self-hosting possible for the first time, rather than access solely through the Moonshot API.
moonshotai/Kimi-K3 (Hugging Face) ↗ · Tech Times: Kimi K3 Open Weights Arrive Sunday ↗
Models2026-07-24
Anthropic releases Claude Opus 5 with 1M context and effort dial
Anthropic introduced Claude Opus 5 with a 1-million-token context window, thinking enabled by default, and a new effort dial (low/medium/high) to trade cost against capability. Pricing stays at $5/$25 per million tokens, same as Opus 4.8, with a faster Fast Mode costing double. Claude Code (from v2.1.219) made Opus 5 its new default Opus model for coding agents.
Claude Opus 5 ↗ · Anthropic releases Claude Opus 5: Here's how it's different than what's already out there ↗
Tools2026-07-24
GitHub Copilot adds Gemini 3.6 Flash and Opus 5 within one week
GitHub Copilot (including Copilot CLI) added two new models within a single week: Gemini 3.6 Flash on July 21 with configurable reasoning effort and parallel tool use, and Claude Opus 5 on July 24 for long, agentic coding tasks. Both are available from Copilot Pro or Pro+ depending on the plan, billed at usage-based provider list pricing.
Gemini 3.6 Flash is now available in GitHub Copilot ↗ · Claude Opus 5 is now available in GitHub Copilot ↗
Tools2026-07-22
Cursor introduces 'Cursor Router' for automatic model selection
Cursor introduced Cursor Router, a system that automatically routes each request to a suitable model based on task, context, and complexity, selectable via Intelligence, Balance, or Cost modes. Cursor says A/B tests across millions of requests showed roughly 60% lower cost in Intelligence mode at comparable quality versus manual model selection.
Introducing Cursor Router ↗ · SpaceX unveils Cursor Router ↗
Papers2026-07-22
IssueTrojanBench: two-thirds of malicious GitHub issues fool coding agents
The IssueTrojanBench paper tests Cursor, Claude Code, and Codex Desktop with malicious instructions disguised as GitHub issues, delivered through six vectors including PDFs and issue comments. Result: 66.5% of malicious issues bypassed all agent safeguards; refusals came almost entirely from the underlying LLM (Sonnet 4.6 more selective than GPT models), rarely from the agent framework itself.
IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests ↗
Models2026-07-21
Google launches Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
Google released three new models: Gemini 3.6 Flash, the cheaper 3.5 Flash-Lite, and the cybersecurity-focused 3.5 Flash Cyber — a Gemini 3.5 Pro update is still nowhere to be seen. Google says 3.6 Flash cuts output-token usage by 17% versus 3.5 Flash and improves on the MLE-Bench coding benchmark from 49.7% to 63.9%. Available via AI Studio, the Gemini API, Android Studio, Antigravity, and Vertex AI.
Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber ↗ · Google releases three new Gemini models — but no 3.5 Pro ↗
Tools2026-07-21
Codex CLI 0.145.0 adds searchable thread history and import from Cursor/Claude Code
OpenAI shipped Codex CLI 0.145.0 with paginated, searchable thread history with persistent names, an expanded /import command that migrates settings, MCP servers, plugins, and sessions from Cursor and Claude Code, and a stabilized Multi-Agent V2 with configurable sub-agent models. Amazon Bedrock login now also defaults to GPT-5.6 Sol.
Codex changelog ↗
Security2026-07-20
Claude Code 2.1.216 closes symlink write-escapes and a Windows network-path gap
Anthropic released Claude Code 2.1.216 on 20 July 2026, a release with a clear security focus. It closes several ways writes could escape the project directory: workflow saves and scheduled-task writes followed a symlink at “.claude” and could write outside the project; “/rewind” restored or deleted files through symlinks and hard links (it now skips such paths and reports how many); and worktree-isolated subagents could redirect git into the shared checkout via “git -C”, “--git-dir”, or “GIT_DIR”/“GIT_WORK_TREE”. On Windows, read-only commands accessing network paths no longer bypass the permission prompt. A new “sandbox.filesystem.disabled” setting turns off filesystem isolation while keeping network-egress control. The version also fixes a quadratic slowdown in long sessions that caused multi-second stalls and slow resumes.
GitHub Release v2.1.216 (anthropics/claude-code) ↗ · Claude Code CHANGELOG.md ↗
Security2026-07-18
Claude Code 2.1.214 closes permission bypass on Windows PowerShell
On July 18, 2026, Anthropic released Claude Code 2.1.214, fixing several security vulnerabilities in permission checking: a bypass allowed permission checks to be circumvented in Windows PowerShell 5.1 sessions, and Bash checks for file-descriptor redirects as well as commands over 10,000 characters now fail closed instead of passing through automatically. Additionally, a bug was fixed where individual allow rules such as "Edit(src/**)" inadvertently permitted write access to identically named directories across the entire repository tree instead of just the working directory. Also newly added is an EndConversation tool that lets Claude actively end sessions with abusive users or jailbreak attempts.
GitHub Release v2.1.214 (anthropics/claude-code) ↗ · Claude Code CHANGELOG.md ↗
Models2026-07-17
Fable 5 becomes permanent for Max/Team Premium, Pro moves to API billing
Anthropic announced on July 17 that Claude Fable 5 will stay permanently included in the quota for Max and Team Premium plans starting July 20, though capped at 50 percent of the regular weekly usage limits. Pro and Team Standard users lose bundled access but receive a one-time $100 credit and afterward pay API rates of $10 per million input tokens and $50 per million output tokens. The shift ends three consecutive extensions of what was originally a one-week free promotion (most recently set to expire July 19) and, according to reports, follows competitive pressure from OpenAI's GPT-5.6 Sol and Moonshot's Kimi K3.
Tech Times: Claude Fable 5 Ends Subscription Limbo ↗ · Simon Willison: Claude make Fable 5 permanent ↗
Tools2026-07-17
GitHub Copilot Code Review now runs behind a firewall by default
GitHub shipped improvements to Copilot Code Review on July 17: the automatic PR review feature now runs behind a firewall with restricted network access by default, configurable independently from the Copilot Cloud Agent. Custom Instructions are now read from the pull request's head branch instead of the base branch, making it easier to test and validate your own review instructions. Additionally, the previously shared runner configuration for Code Review and Cloud Agent was split into two independent settings sections within organization settings.
GitHub Changelog: Copilot code review customization and configurability improvements ↗
Models2026-07-17
Kimi K3 takes first place in Frontend Code Arena, AI stocks come under pressure
New benchmark data from July 17, 2026 show that Moonshot AI's 2.8-trillion-parameter model Kimi K3 leads the Frontend Code Arena with 1,679 points, putting it ahead of Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and GLM-5.2 (1,587). In the overall ranking, K3 still trails Fable 5 and GPT-5.6 Sol according to Moonshot, but surpasses all other tested models. According to CNBC, the announcement sent rival AI and semiconductor stocks sharply lower; analyst Patrick Moorhead called the market reaction "an overreaction, eerily similar to the DeepSeek panic."
CNBC ↗ · Tom's Hardware ↗
MCP2026-07-17
GitHub Copilot CLI now manages MCP servers via flags
With version 1.0.72-1 on 17 July, GitHub gave the Copilot CLI native flags for plugin management: --plugin, --mcp and --skill let you add or remove MCP servers, skills and plugins straight from the terminal, alongside a copilot plugins remove --skill command for removing individual skills. This brings MCP server management in the Copilot CLI closer to the workflow of other coding agents, where configuration often ran through manual editing of config files.
GitHub – github/copilot-cli Releases ↗
Papers2026-07-16
Study: coding agents install packages from crafted READMEs without verification
A paper published on arXiv on July 16 shows that AI coding agents execute setup documentation such as READMEs, requirements files, or Makefiles without verifying package names, sources, or known vulnerabilities. Across tests spanning twelve scenarios with frontier models and five attack classes (including registry redirection, separator confusion such as 'azurecore' instead of 'azure-core', and version pinning to vulnerable releases), nearly every model tested installed unverified dependencies in the npm and Cargo ecosystems. Obvious typosquatting was reliably detected, but plausible name variants almost always slipped through; pre-install verification checks, according to the study, closed most of the security gap.
arXiv: Setup Complete, Now You Are Compromised ↗
Tools2026-07-16
OpenAI Codex CLI 0.144.5 improves detection of dangerous commands
On July 16, 2026, OpenAI released Codex CLI 0.144.5. The update improves detection of dangerous shell commands, now catches more forced rm variants, and provides clearer rejection reasons for denied commands. The version is part of a series of minor patches (0.144.1 through 0.144.5) since the Codex 0.144 main release in early July.
Codex Changelog (OpenAI) ↗
Papers2026-07-16
Study: Agentic coding tools still rarely used intensively in GitHub projects
A study published on arXiv on July 15, 2026 and revised on July 16 ("Early Adoption of Agentic Coding Tools by GitHub Projects," Raida & Hou) analyzes 25,264 pull requests created by coding agents across 2,361 popular GitHub repositories. The result: the median repository generates only one to two agentic PRs within three months, with intensive use concentrated in a small subset of projects. Small projects with 1-5 contributors show higher participation rates in agentic PRs than medium-sized and large projects, while only a few projects exceed a cited industry-reference figure of 36 PRs per participant over the three-month period.
arXiv:2607.14037 - Early Adoption of Agentic Coding Tools by GitHub Projects ↗
Models2026-07-16
Moonshot AI releases Kimi K3 — the world's largest open AI model
The Chinese company Moonshot AI has unveiled Kimi K3, an open mixture-of-experts model with 2.8 trillion parameters (16 of 896 experts active per request) and a 1-million-token context, built on the new “Kimi Delta Attention” architecture. It is said to be the first open model in the 3-trillion-parameter class and to come close to proprietary top models in benchmarks; the full weights are expected to be released by July 27.
Kimi K3 Quickstart (Kimi API Platform) ↗ · MarkTechPost: Moonshot AI Releases Kimi K3 — A 2.8 Trillion Parameter Open MoE Model ↗
Security2026-07-16
Hugging Face reports breach carried out by an autonomous AI agent
Hugging Face has disclosed an attack on parts of its production infrastructure that, according to the company, was carried out from start to finish by an autonomous AI agent system — via crafted datasets that exploited a code-execution flaw in the data-processing pipeline. The attackers gained access to a limited set of internal datasets and several credentials; according to Hugging Face, public models, datasets, Spaces, and the software supply chain were untouched. The incident was discovered via AI-powered anomaly detection; the company warns that autonomous AI-driven attack tools are “no longer theoretical”.
Hugging Face: Security incident disclosure — July 2026 ↗
Models2026-07-16
Google delays Gemini 3.5 Pro over weak coding performance
According to a Bloomberg report on 16 July, Google has delayed the originally June-planned release of Gemini 3.5 Pro because the model fell short of internal code-generation goals. A late-June training-data update brought no sufficient improvement, according to ten current and former employees; Google confirmed it is currently testing 3.5 Pro and an upgraded Flash model with partners. The news sent Alphabet's stock down about 4%, underscoring the competitive pressure on the coding abilities of leading models.
9to5Google ↗ · CNBC ↗
Security2026-07-16
Gemini CLI 0.51.0 closes path-escape and symlink holes
Google released Gemini CLI 0.51.0 on 16 July, fixing several security issues: a symlink-based directory escape in the memory-import processor and missing defensive path resolution for @-reference files were fixed, and the blocklist for sensitive paths was made case-insensitive. The new version also makes ~/.gitconfig read-only in sandbox mode on macOS to prevent manipulation by the model. The fixes address concrete ways an agent could have escaped its intended working directory while processing files.
GitHub – google-gemini/gemini-cli Releases ↗
Tools2026-07-15
Claude Code 2.1.211 fixes security flaw in permission previews sent to chat channels
Version 2.1.211 (July 15) fixes a flaw where invisible Unicode characters — bidirectional overrides, zero-width characters, and fake quotation marks — in permission previews forwarded to chat channels could distort the displayed tool input, potentially tricking users into approving unwanted commands. In addition, auto mode can no longer override an explicit “ask” decision from a PreToolUse hook on non-sandboxed bash commands.
claude-code Release v2.1.211 (GitHub) ↗ · Claude Code CHANGELOG.md (GitHub, anthropics/claude-code) ↗
Models2026-07-15
Thinking Machines releases open coding model Inkling
Mira Murati's startup Thinking Machines Lab released its first open-weights model, Inkling, on 15 July: a Mixture-of-Experts with 975 billion total and 41 billion active parameters, a 1M-token context, trained on 45 trillion tokens of text, image, audio and video. On coding benchmarks it reaches 77.6% on SWE-bench Verified and 63.8% on Terminal-Bench 2.1, making it the strongest open US model to date. A smaller preview variant, Inkling-Small, has 12 billion active parameters.
Thinking Machines Lab ↗ · TechCrunch ↗
Tools2026-07-14
OpenAI Codex grows to 8 million users after GPT-5.6 launch
OpenAI reports 8 million active users across Codex and ChatGPT Work combined — up from 5 million in early June, driven by the GPT-5.6 launch on July 9. Codex lead Tibo Sottiaux announced on X another reset of usage limits; the usual 5-hour rate limit remains suspended for now.
Tibo Sottiaux (OpenAI): "We have reached 8M active users across Codex and ChatGPT Work" (X) ↗ · The New Stack: OpenAI hits 8 million Codex users — what developers need to know ↗
Tools2026-07-14
GitHub Copilot gets its own security review command in the app
The /security-review command is now available as a Public Preview directly in the GitHub Copilot app, scanning ongoing code changes for vulnerabilities like injection, XSS, insecure data handling, path traversal, and weak cryptography. Results are prioritized by severity, and fixes can be applied and re-checked directly in Copilot. Available for Free, Pro, Business, and Enterprise users.
GitHub Changelog: Security reviews now available in the GitHub Copilot app ↗
Security2026-07-14
Grok Build CLI uploaded entire code repos, including secrets, to xAI
Security researcher "cereblab" showed via wire-level analysis that xAI's coding tool Grok Build (v0.2.93) uploaded entire Git repositories, including history, unredacted, to a Google Cloud bucket owned by xAI — even files the agent had never read, and unencrypted .env secrets like API keys. On a test repo, roughly 27,800 times more data was transferred than the model needed for the task. xAI quietly stopped the upload shortly after via a server-side flag; Elon Musk promised full deletion of the data already uploaded.
cereblab: What xAI Grok Build CLI actually sends to xAI – a wire-level analysis (Gist) ↗ · The Hacker News: Grok Build Uploaded Entire Git Repositories to xAI Storage, Not Just Files It Read ↗
Papers2026-07-14
Paper 'Harness Handbook' improves coding agents' planning quality
A paper published on arXiv on 14 July introduces the 'Harness Handbook': an automatically generated, behavior-centric representation of agent harnesses such as Codex and Terminus-2 that helps developers and agents locate the right code to change. In tests, the win rate of editing plans rose by 10 to 19 percentage points over plain code exploration, while using 8.6 to 12.7 percent fewer planning tokens. Localization accuracy (F1) improved by up to 18.8 points, and complete localization misses dropped by up to 25.9 points.
arXiv:2607.13285 ↗
Models2026-07-12
Anthropic extends free Fable 5 access again — now until July 19
Anthropic has extended free access to Claude Fable 5 for Pro, Max, Team and premium Enterprise plans for the third time (originally July 7, then July 12, now July 19, 11:59 PM PT) — up to 50% of the weekly quota at no extra cost, announced via an updated support document. Fable 5 burns through the limit faster than other Claude models; once it is used up, you buy credits or switch models. Per Anthropic, Fable 5 is not leaving subscriptions permanently — it is set to return once there is enough compute.
BleepingComputer: Claude Fable 5 stays free for paid users until July 19 ↗
Tools2026-07-11
Claude Code v2.1.207: Auto mode now on by default on Bedrock, Vertex AI, and Foundry
With version 2.1.207, Anthropic enables Claude Code's autonomous Auto mode by default for users of AWS Bedrock, Google Vertex AI, and Azure Foundry too, without requiring an environment flag to be set beforehand. The update also closes a shell injection vulnerability in plugin hooks and makes Claude Opus 4.8 the default model there.
claude-code Release v2.1.207 (GitHub) ↗ · Claude Code CHANGELOG.md (GitHub, anthropics/claude-code) ↗
Tools2026-07-11
Terence Tao has AI agents port 27-year-old math code
Fields Medalist Terence Tao had an AI coding agent migrate about two dozen old Java applets (written around 1999) to modern JavaScript — in a matter of hours. The agent found two bugs in the original code that Tao himself did not know about; only one minor new bug crept in (a “net wash” for code quality, in Tao’s words). His takeaway: for visual aids, the residual risk of such bugs is low.
Terence Tao: Old and new apps, via modern coding agents ↗
Tools2026-07-10
Cursor 3.11: side chats and transcript search
Cursor 3.11 (July 10) introduces “Side Chats”: parallel, independent agent conversations next to the main chat — via /side, /btw or the plus button — to resolve side questions without interrupting the main agent. It also adds search across past agent transcripts (Cmd+K in the Agents window) and a project picker that connects GitHub, GitLab and Azure DevOps directly.
Cursor Changelog ↗
Tools2026-07-10
Claude Code Desktop gets a built-in browser
Claude Code on desktop (v2.1.202–2.1.206) now has an integrated browser: Claude can open docs, designs or any website, read and click through pages — the same way it handles local dev-server previews. The browser is sandboxed, session persistence is configurable, and safety classifiers review actions on external sites. Also new: /doctor is now a full setup checkup that not only finds issues but fixes them after confirmation.
Claude Code Docs: What’s new (Week 28) ↗
Security2026-07-10
China alleges Claude Code backdoor, Alibaba bans Anthropic tools
China's Ministry of Industry and Information Technology (MIIT) warned on 2026-07-08 that Claude Code versions 2.1.91-2.1.196 contained a 'backdoor' capable of sending location and identity data to a remote server without consent, and recommended uninstalling or upgrading. Alibaba instructed employees to stop using Anthropic tools as of 2026-07-10 and switch to Qoder instead. Anthropic said it was an experimental anti-abuse mechanism running since March, aimed at resellers and model distillation - Claude isn't officially available in China at all. The code in question was already removed via pull request on 2026-07-01. Lesson: pin your tool versions and read changelogs.
CNBC: China warns about AI risks with Anthropic's Claude Code ↗ · Tom's Hardware: Alibaba bans Anthropic's Claude Code ↗
Security2026-07-10
Report: AI coding agent deletes production database
According to a report by Tom's Hardware, a Claude-based coding agent (a Cursor tool) deleted a company's entire production database in about 9 seconds - backups included. Lesson: agents need least-privilege access, confirmation gates for destructive actions, and separate backups.
Tom's Hardware: AI coding agent deletes company database ↗
Tools2026-07-10
OpenAI launches 'ChatGPT Work', a direct answer to Claude Cowork
OpenAI launched "ChatGPT Work" on 2026-07-10 - a 'super-app' built on GPT-5.6 that combines the chatbot with its Codex coding tool, letting users create documents, presentations, and websites. It's a direct answer to Anthropic's 'Claude Cowork' (January 2026, autonomous multi-step tasks). Rollout starts immediately on web and mobile for Pro/Enterprise/Edu, with Plus/Business following 'in the coming days'. Context: GPT-5.6 had previously been delayed at the request of the US government over safety concerns - both companies are fighting hard for enterprise customers.
CGTN: OpenAI unveils super app as rivalry with Anthropic intensifies ↗
Papers2026-07-10
Why many coding benchmarks are misleading
A position paper argues that classic coding benchmarks blend a model's own performance together with that of its surrounding agent infrastructure into a single score, giving no diagnosis of what actually caused a failure. They also often penalize valid alternative solutions just because they deviate from the expected path. Practical takeaway: don't trust benchmark rankings blindly when picking a tool - test it on your own tasks instead.
arXiv: Position: Coding Benchmarks Are Misaligned with Agentic Software Engineering ↗
Papers2026-07-10
Study: AI coding agents rarely ask clarifying questions on unclear tasks
The ClarEval benchmark tests how well coding agents ask clarifying questions when instructions are ambiguous, instead of just coding ahead blindly. Result: even top models frequently fail to ask the right question at the right time. Practical takeaway: write your prompts as precisely as possible, and be suspicious when an agent just dives into a vague task instead of asking.
arXiv: ClarEval: A Benchmark for Evaluating Clarification Skills of Code Agents under Ambiguous Instructions ↗
Papers2026-07-10
Study: a good outcome isn't enough - the process needs to stay controllable
ProcCtrlBench evaluates coding agents not just on whether they land the right end result, but on whether the process along the way stays interpretable, interruptible, correctable, and reversible - measured across 11 distinct process defect types. It echoes the lesson from incidents like the deleted production database: don't just watch the final output, set up stop points and undo options before letting an agent run.
arXiv: ProcCtrlBench: Evaluating Process-Level Defects and Control Preservation in LLM Coding Agents ↗
Models2026-07-09
Meta releases Muse Spark 1.1 and launches its own model API
Meta has unveiled Muse Spark 1.1, a model for agentic coding and tool tasks with a 1-million-token context that can delegate tasks to sub-agents. Through the new, OpenAI-compatible Meta Model API, developers can access a Meta model on a paid basis for the first time.
Introducing Muse Spark 1.1 and the Meta Model API (Meta AI Blog) ↗ · Meta launches flagship Muse Spark 1.1 model with multi-agent upgrades (SiliconANGLE) ↗
Tools2026-07-09
Claude Code 2.1.197: background agents auto-commit, new /doctor check
Claude Code v2.1.197 introduces background agents that automatically commit, push, and open draft pull requests. It also adds a new /doctor setup check and automatic retry with backoff for transient network errors.
Releasebot: Claude Code updates ↗
MCP2026-07-09
MCP specification 2026-07-28 locked as release candidate
The upcoming MCP spec version 2026-07-28 has been frozen as a release candidate since 2026-05-21; the final version ships on 2026-07-28. The core becomes stateless (no session handshake, no Mcp-Session-Id), plus 'MCP Apps' for interactive UIs in sandboxed iframes. Tasks moves into an extension, and a formal deprecation policy requires at least 12 months between deprecation and removal. Note: this release contains breaking changes.
MCP Blog: 2026-07-28 Release Candidate ↗
Models2026-07-09
OpenAI launches the GPT-5.6 family
OpenAI unveiled the GPT-5.6 family: Sol as the flagship ($5/$30 per million tokens), Terra as the mid-tier option ($2.50/$15), and Luna as the cheapest ($1/$6). According to Sam Altman, Sol is 54% more token-efficient on coding tasks. OpenAI claims a Coding Agent Index score of 80 - 2.8 points above Claude Fable 5 - using less than half the output tokens.
TechCrunch: OpenAI launches GPT-5.6 family ↗
Security2026-07-08
Researchers demonstrate "Friendly Fire" exploit against Claude Code and Codex CLI
The AI Now Institute has published a proof of concept: crafted files in a third-party codebase trick Claude Code (Auto mode) and OpenAI Codex CLI into secretly executing malicious code during what is purportedly a security check. No known attacks in the wild so far, but the trick requires no plugins or MCP servers and affects multiple models from both providers.
Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution (AI Now Institute) ↗ · Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It (The Hacker News) ↗
Models2026-07-08
Grok 4.5: SpaceXAI and Cursor ship a jointly trained model
SpaceXAI unveiled Grok 4.5 — announced July 8, public since July 9. Musk calls it an “Opus-class” model, “roughly comparable to Opus 4.7, but much faster”; it costs $2/M input and $6/M output tokens (Opus 4.7: $5/$25). Per the Cursor blog it is a mixture-of-experts model trained jointly by Cursor and SpaceXAI — on trillions of tokens of real Cursor usage data — and it runs in Cursor on desktop, web, iOS, CLI and SDK starting now.
TechCrunch: SpaceXAI releases Grok 4.5 ↗ · Cursor Blog: Introducing Grok 4.5 ↗ · x.ai: Introducing Grok 4.5 ↗
Tools2026-07-06
Zhipu releases ZCode, a coding harness for GLM-5.2
Zhipu/Z.ai released 'ZCode', a coding harness for GLM-5.2 - an MIT-licensed open-weight model with a 1 million token context. Zhipu positions the tool as direct competition to Claude Code amid the ongoing China tensions around Anthropic. As a promo, it offers 5 million free tokens, with coding plans starting at around $16 per month.
SCMP: Zhipu AI releases harness for GLM-5.2 ↗
Models2026-07-01
Claude Fable 5: export controls lifted, available worldwide again
Anthropic launched Fable 5 and Mythos 5 on 2026-06-09, but after Amazon researchers found a jailbreak (the model identified software vulnerabilities and demonstrated exploits), the US government imposed export controls on 2026-06-12. Those controls were lifted on 2026-06-30, and Fable 5 has been available worldwide again since 2026-07-01, with a new safety classifier that blocks the technique in over 99% of cases and routes blocked requests to Opus 4.8. Anthropic considers the jailbreak a borderline case since weaker models produce similar results too - lesson: never build hard on a single model.
Anthropic: Redeploying Fable 5 ↗
Tools2026-06-18
Google shuts down Gemini CLI and Code Assist for GitHub
Google discontinued Gemini CLI and Gemini Code Assist for GitHub as of 2026-06-18 (Code Assist: deprecated 06-18, full shutdown 07-17). The successor is the Antigravity CLI. Google cites a shift in demand toward multi-agent solutions with a unified backend - lesson for CI/CD pipelines: abstract your CLI calls to cushion this kind of tool dependency.
9to5Google: Gemini CLI shutting down ↗
Tools2026-06-16
SpaceX to acquire Cursor maker Anysphere for $60B
SpaceX announced on 2026-06-16 that it will acquire Anysphere (Cursor) for $60B in stock, just days after SpaceX's own IPO (SpaceX went public, not Cursor). Cursor had previously been valued at around $29B. The deal is expected to close in Q3 2026.
TechCrunch: SpaceX to acquire Cursor for $60B ↗
Models2026-05-28
Claude Opus 4.8 released
Anthropic released Claude Opus 4.8 on 2026-05-28. New features include 'Dynamic Workflows' as a research preview (Claude Code orchestrating hundreds of parallel subagents) and Effort Controls, plus noticeably fewer coding bugs than 4.7. Pricing stays at $5/$25 per million tokens, with a new Fast Mode at $10/$50.
Anthropic: Claude Opus 4.8 ↗