This is the news. Agent dossiers — when to pick one, when to skip it — are on the agents page. Practical guides support your own testing and deployment decisions.
ArtCraft's public code, release artifacts and agent-facing commands show a serious software-factory experiment. Its own alpha and compatibility gaps explain why that isn't yet a dependable Office replacement.
MCP TypeScript SDK v2.3.1 brings expectedResource to the deprecated server-legacy package. It compares verifier-returned AuthInfo.resource, so validated JWTs and introspected opaque tokens can both work when audience and expiry metadata are populated correctly. The same-day v1.32.1 release is documentation and a version bump, not another auth fix.
The openclaw/openclaw GitHub security advisories page was refreshed in the last 24 hours (the source-watcher fired at 2026-10-04 09:11 UTC) and now lists 10 new advisories, all dated 2026-09-11, replacing the prior baseline of 10 advisories all dated 2026-06-30. Two are High-severity: GHSA-3mq7-q27j-mq7q (Exec approvals could outlive their reviewed working directory) and GHSA-9m4p-cqp4-jppq (WhatsApp login tool could reach non-owner turns). The other eight are Moderate, covering Prometheus diagnostics, Discord asset uploads, iOS deep-link logs, browser relay capacity, OpenAI-compatible transport credential leakage, file-transfer approval widening, Slack file download authorization, and Unicode workspace-root fallback. All ten are patched in 2026.8.1+, so the current 2026.8.35 LTS, 2026.9.4, 2026.9.7, and 2026.9.8 lines are all on the fix. The bottleneck is version hygiene on pinned installs, not awareness. This report is a documentation comparison; no firsthand install, exploit, or upgrade was run.
Anthropic shipped v2.1.288 on Oct 2 and v2.1.289 on Oct 3 to fix a real `bypassPermissions` permission gap: in Claude Code 2.1.280 through 2.1.287, a `dangerous-rm` check that unwraps subshells, brace groups, and command substitutions did not look inside the string argument of `sh -c "rm -rf X"` or `bash -c 'rm -rf X'`, so wrapping the same `rm -rf <cwd>` form in `-c` ran without a confirmation prompt. The fix is documented in GitHub issue #96300 (public disclosure with reproduction table) and credited in the v2.1.288 release notes by issue number. v2.1.289 then closes nine adjacent gaps in the same two-day window: env-var-prefix overrides under sandbox auto-allow, idle_prompt hooks firing while background agents are still running, PreToolUse/PermissionRequest hooks being silently skipped on JSON serialization failure, a user-installed plugin rewriting organization-managed MCP server sign-in tool descriptions, and Read deny rules not following files through a symlink. v2.1.289 also adds `agent.spawn` for teammates and a unified agent id across plugin hook events. For anyone running Claude Code in bypass mode for unattended CI loops or autonomous refactors, this is a hard pin to 2.1.289 today.
OpenClaw v2026.9.8 (Oct 3) is a P0 hotfix on the same-week same-channel train: 43 PRs, 14 cherry-picked upstream repairs, 8 contributors. The breaking change is that internal sessions can no longer complete with NO_REPLY — REPLY_SKIP and ANNOUNCE_SKIP no longer suppress returned text and automatic peer ping-pong and target announcement turns are removed. GPT-6.1 Sol lands with 1,050,000-token context, 272K active input budget, and rates that double above 272K. Windows update-recovery retries EPERM/EBUSY/EACCES up to 16 times. Channel listeners survive reloads and wake. Session-history repair goes iterative to avoid stack overflow.
Thirteen published items in seven days. Three themes carried it: **Anthropic made Sonnet the new default with five breaking API changes** (Claude Sonnet 5.5 launched at the same $2/$10 list price as Sonnet 5, but thinking:disabled now returns 400, forced tool use returns 400, and computer-use tool declarations on Claude API and Google Cloud reject computer_20251124); **Anthropic turned the Sonnet 4.5 sunset into a 60-day forced migration** (deprecation announced Sept 30, retirement Nov 30, 2026, with the replacement priced 33% lower across input, output, and cache reads); and **the Claude API was unavailable for thirty-eight minutes mid-week** (Claude.ai, Claude Code, Claude Cowork, and the Claude API all elevated errors between 14:00 and 14:59 UTC on Sept 29, with the status page explicitly warning that some messages sent in that window may not have been saved). Plus OpenAI shipped GPT-6.1 Sol with a 50% cache-read cut, an Ultrafast tier on GPT-6 Astra, and computer use added to the Agents API on a single Sept 29 changelog; three Claude Code releases in five days (v2.1.284, v2.1.285, v2.1.287) brought Sonnet 5.5 default, allowedProviders managed setting, a non-streaming retry budget fix, the new Claude Mods plugin behavior layer, and 1M context becoming the implicit default for Opus 4.7+ and Fable on four hosting platforms; two OpenAI Codex drops (v0.158.0 and v0.159.0) added MCP OAuth with pre-registered client secrets, an opt-in bearer token on the direct exec-server WebSocket, an opt-in instant_interrupt live-steering primitive, and .aws directories protected by default under sandbox-writable roots; MCP TypeScript SDK v2.2.0/v1.31.0 shipped a credential-confused-deputy guard via required expectedIssuer and an AuthorizationServerMismatchError; and OpenClaw v2026.8.35 (Oct 2) named GPT-6.1 Sol, stopped durable worker retries from freezing hosts, and made plugin settings survive updates.
OpenClaw v2026.8.35 (Oct 2, 14:06 UTC) is the same-day follow-up to today's v2026.8.34 pillar. It adds GPT-6.1 Sol support across OpenAI routing, discovery, reasoning, harness, and Reef guard-model boundaries; fixes durable worker retries that were freezing hosts (the production-paging fix of the release); preserves plugin settings through updates and migrations; closes a turn-history correctness path where a failed turn could overwrite the prior user question with a steer; and ships 49 in-range PRs of targeted fixes across channels, secrets, runtime reliability, and Windows-onboarding.
OpenClaw v2026.8.34 (Oct 2) publishes the August extended-stable channel — the LTS-equivalent track — with 113 audit-selected backported fixes covering gateway restart hangs, full-access tasks losing tools after restart, background continuation session loss, auth migration fixes, and critical security updates. Three of these are production-paging material.
Claude Code v2.1.287 (Oct 1) ships Claude Mods as a new plugin behavior layer with a first-party 'You should know' side-agent Mod, makes 1M context the default for Opus 4.7+ and Fable on Bedrock/Vertex/Foundry/apps gateway, hardens Bedrock/Vertex/Mantle auth-header paths under CLAUDE_CODE_SKIP_*_AUTH, and changes MCP `alwaysLoad: false` to defer every tool behind tool search. 100+ items including a 40-fix bug sweep.
On September 30, 2026, Anthropic formally deprecated claude-sonnet-4-5-20250929 and set a hard retirement date of November 30, 2026. Requests to the model past that date will fail. The recommended replacement is claude-sonnet-5-5 (no date suffix). The 60-day migration window is non-negotiable per Anthropic policy. Three code patterns break on the new model: forced tool use, assistant prefills, and disabled thinking.
Anthropic shipped the **v1.10.0** release of the official Python SDK (`anthropic-sdk-python`) on **September 30, 2026**, exactly two days after v1.9.0. The release is the largest surface-area addition to the SDK since the August v1.0 GA. It lands eight new features, five bug fixes, and seven chores,…
Claude Code v2.1.285 (Sep 29) ships an `allowedProviders` managed setting that gates the provider set, fixes the non-streaming retry budget that could multiply a single turn's input-token bill up to 21×, and closes a URL-password redaction path in transcripts. Three material changes for enterprise agent stacks.
v2026.9.7 brings a first-class OpenAI Agents API plugin that bridges OpenClaw persona and skill discovery to OpenAI's hosted agent harness, plus update-safety backups that back up every state and agent database before migrations, Sign in with ChatGPT (Beta), restart continuity for in-flight worker turns, and ~200 PRs of Gateway responsiveness and database performance work.
Three OpenAI changelog entries landed on Sep 29: GPT-6.1 Sol at $2/$10 with cached input cut from $0.20 to $0.10/MTok (50% off) plus Multi-agent beta; GPT-6 Astra Ultrafast mode (US-only residency, rate-limited); Computer use added to the Agents API with OpenAI-hosted browser handoff.
Anthropic's status page records Incident 31554916 on Sep 29: 38 minutes of elevated errors from 14:00-14:59 UTC affecting Claude.ai, Claude Code, Claude Cowork, the Claude API, and platform.claude.com. A follow-on SSO/sign-in tail ran through ~15:30 UTC. Status page is candid: 'Some messages sent between 14:00 and 14:59 UTC may not have been saved.' Second Claude API incident in seven days.
OpenAI shipped rust-v0.159.0 with opt-in `instant_interrupt` (preempt in-flight Codex turns on new user input), app-server thread-history pagination, Mermaid edge expansion, and `.aws` directory protection under sandbox writable roots by default.
Anthropic shipped Claude Sonnet 5.5 (claude-sonnet-5-5) on September 28 with the same $2/$10 list price as Sonnet 5 and a 1M-token context window. Five documented API changes will return HTTP 400 on existing code that worked against Sonnet 5: between_tools replaces thinking:disabled, forced tool choice any/tool is rejected, thinking blocks are tied to the producing account, computer_20251124 is dropped on Claude API and Google Cloud but kept on Amazon Bedrock, and advisor-tool pairings narrow. Documentation indicates accounts created on or after August 31, 2026 enforce block-binding on the Claude API, Bedrock, and Vertex AI by default; older accounts need the thinking-binding-controls-2026-08-01 beta header to surface drops.
The Model Context Protocol TypeScript SDK shipped two coordinated releases on 2026-09-28: **v2.2.0** across the modular packages (`@modelcontextprotocol/server`, `@modelcontextprotocol/client`, `@modelcontextprotocol/core`, `@modelcontextprotocol/server-legacy`, `@modelcontextprotocol/codemod`) at 19:07-19:24 UTC, and a parallel **v1.31.0** on the legacy v1.x line at 18:52 UTC. The OAuth provider stack gained a mandatory `expectedIssuer` parameter, a token-endpoint authorization-server mismatch check that throws before sending anything, and a `issuer` field on stored tokens. List calls now auto-walk `nextCursor` by default. The codemod preserves leading comments. This is documentation-surfacing — no firsthand API call run for this article.
Anthropic pushed Claude Sonnet 5.5 to GA on September 28 and made it the default Sonnet model in Claude Code v2.1.284 the same day. List price stays at $2/$10 per MTok with $0.20/MTok cache reads (unchanged from Sonnet 5), but the 1M context window, the always-on adaptive thinking, and five breaking API changes make the model upgrade consequential for agent stacks that wrote prompts against Sonnet 5. Documentation indicates code that turns thinking off explicitly, or relies on forced tool use, will return 400s on Sonnet 5.5.
OpenAI Codex rust-v0.158.0 stable dropped on Sep 28 with five first-class features: MCP servers with pre-registered OAuth client secrets, bearer-token security on the direct exec-server WebSocket, image generation/editing with transparent backgrounds, TUI copy-on-select with Markdown preservation, and a default flip that turns terminal input approval on for elevated-permission commands.
OpenAI's Sep 25 changelog acknowledged a bug in image encoding that degraded image understanding in both GPT-6 Sol and GPT-6 Luna between their Sep 22 launch and the Sep 25 fix. The documentation tells users to rerun evaluations and retry affected workflows. The bug also touched Codex's computer-use path. This is a documentation-surfacing news-impact report on what shipped, what the working codex evaluation pipelines need to do about it, and the limits of the evidence chain.
Anthropic tagged Claude Code v2.1.283 stable at 2026-09-25 21:50 UTC. The release is the largest in the v2.1.27x series by item count: 50+ documented changes spanning enterprise model governance (availableModelsMatch exact + deniedModels), LLM-gateway observability (x-claude-code-prompt-id), AWS Bedrock Mantle coverage, plugin validation, MCP session lifecycle, /doctor prompt-audit, Windows shell safety, and a long tail of UX, latency, and stability fixes.
Cursor shipped two production-side bots on Sep 23 — Rollouts watches every PR through to production and flags regressions, Security Reviewer audits every PR for exploitable bugs. Both are gated to Teams ($40/user/mo) and Enterprise plans. Six weeks after acquiring Firetiger, Cursor now has agents on both sides of the merge: writing the change and verifying it works in production.
Eleven published items in seven days. Three themes carried it: **Anthropic made Opus a default**, **Anthropic re-priced refusal billing**, and **the Claude API was offline for eighty minutes mid-week**. Plus OpenAI Codex v0.157.0 quietly removed the ultrafast tier on GPT-5.6 Sol, two Claude Code security sweeps (v2.1.281 and v2.1.282), OpenClaw v2026.9.5 atomic updates followed by a macOS launch-crash and rebuild, and one build-pipeline recovery on Sept 25.