AI Developer News: Week Ending October 23, 2026

This week’s AI developer news is dominated by two themes: AI models are now finding real CVEs, and everyone is racing to give agents real money. PHP shipped emergency patches for RCEs that frontier models helped discover, Anthropic launched a proactive agent app and its first new API version since 2023 (with a breaking change you’ll want to know about), OpenAI took its Decisions API generally available, Stripe gave AI agents bank accounts, and Meta’s Muse agent passed 10 million users on the strength of shopping. Here’s what matters for builders, with links to the original announcements.

All facts verified October 23, 2026. We previously covered the PHP2Shell patches from the WordPress angle — here’s that patch guide; this week we cover the AI angle.

Table of Contents

PHP2Shell: frontier AI models helped find a pair of critical PHP RCEs

The headline security story of the week was discovered, in part, by the very models you build on. On October 21, PHP shipped security-only releases — 8.5.12, 8.4.27, 8.3.36, and 8.2.35 — fixing two critical memory-safety RCEs in the engine core: CVE-2026-109618 (out-of-memory corruption via longjmp) and CVE-2026-109619 (stale global state / use-after-free), both rated CVSS v4.0 9.2 and exploitable by unauthenticated requests, with mod_php deployments at highest risk.

Wiz Research, which nicknamed the pair PHP2Shell, found the first bug on August 19 and reported it to PHP on August 24. The notable part for developers: Wiz used its Wiz Atlas tool plus GPT-5.6 Sol and Gemini 4 Argon to develop working exploits against current WordPress, Drupal, and Joomla container images. The flaws affect every PHP version back to 5.4. PHP says updating is the only complete mitigation, with minimal source backports provided for end-of-life 7.4, 8.0, and 8.1. Public exploit tooling appeared the same day as the patches.

Why it matters: this is the first major CVE pair discovered and exploited with frontier AI assistance, which proves AI-accelerated offensive security research is real and already here. Expect coordinated-disclosure processes — and your own threat models — to start assuming attackers develop exploits with AI help. And if you run PHP anywhere, patch first and philosophise later.

OpenAI’s Decisions API goes generally available, powered by GPT-6 Luna

Announced October 22 on OpenAI’s Developer Community forum, the Decisions API is out of beta and generally available, powered by GPT-6 Luna (unveiled at OpenAI’s developer day on September 29). It returns typed, bounded outputs — a probability, a choice from options, or a rubric score — for classify/route/score calls, which OpenAI claims run up to 10x faster than GPT-6 Luna via the Responses API, with response times of 100–180ms.

The pricing is the real story: $0.10 per million non-cached input tokens, $0.01 per million cached input, and output tokens are not charged at all. GA adds native GPT-Live integration, prompt caching, native JSON inputs, image instructions per question, and up to 400 questions per request. Early trials ran with Box, Datadog, Ramp, and Speak — Speak cut language-tutor action lag from 1.4s to roughly 0.13s. Inference routes geographically across US and Europe regions, with zero-data-retention available for eligible customers.

Why it matters: if your agents make lots of classify/route/pick calls, you’ve been paying full generation prices for bounded decisions. Audit those call sites and swap them — the 10x speed claim plus free outputs make this the cheapest way to get model judgement into hot paths.

Anthropic launches Claude Duet, a proactive agent app (waitlist)

Anthropic announced Claude Duet on October 21: a standalone iOS/Android/claude.ai app where Claude works as a proactive agent rather than a chatbot. It delivers morning briefs, monitors connected apps every few hours, surfaces items that need your attention, and keeps working when your devices are off — it runs on its own VM.

It’s built on Claude Opus 5.5 and integrates with 6,000+ Claude Marketplace connectors (Gmail, Slack, Stripe, QuickBooks). Purchases run through Stripe Link‘s digital wallet (virtual cards, with the user approving each spend) and Shop Pay. It’s US-only, for Pro and Max subscribers, behind a waitlist as a research preview; background checks count toward existing plan usage limits, and Anthropic warns heavy background use can burn through them “sooner than with chat alone.” Duet asks before sending messages, making payments, signing, or deleting, and Anthropic openly acknowledges it can “misread a thread or miscalibrate what’s urgent,” with a classifier and safety system prompt guarding against sensitive-info leaks.

Why it matters: the agent-platform war — OpenAI’s Dots, Meta’s Muse, now Claude Duet — is converging on the same shape: connectors, background VMs, and approval-gated payments. That’s the shape of consumer agents to build against, and the Stripe Link / Shop Pay plumbing shows where agentic-commerce standards are settling.

Claude API version 2026-10-21: Anthropic’s first new API version since 2023 — and it breaks sampling controls

Anthropic shipped its first new official API version since 2023 on October 21 (2023-06-01 stays available, not deprecated). New major SDKs — Python 2.0.0, plus TypeScript, Go, Java, C#, PHP, Ruby — and ant CLI 2.0.0 default to it; you select it via the anthropic-version header.

The breaking change: temperature, top_p, and top_k are removed from the Messages API and Message Batches. Passing them — even at their default values — now returns a 400. Omit them and requests sample at defaults (temperature 1.0), matching Claude 4.7+ behavior. There is no replacement setting; Anthropic recommends prompting as the substitute. The new version also adds 400s for document blocks directly in assistant turns and non-empty citations in user turns, and renames some deprecated Admin and Enterprise Analytics API response fields. Anthropic says versions will now ship more frequently so developers can handle breaking changes before new models require them.

Why it matters: pin your SDKs and test before upgrading. Anyone relying on custom sampling — low-temperature classification pipelines, for example — has real migration work, and the new versioning cadence means API behavior will churn faster going forward.

Claude Code 2.1.307: an OpenTelemetry signal that shows which guardrail blocked your agent

Claude Code’s October 21 release added a genuinely useful observability feature: a new OTel event, claude_code.safety_event, that fires for every tool call a permission rule, auto mode, safety check, hook, or sandbox stopped, prompted, or sent back. Instead of grepping session logs to figure out why an agent froze, you can now see exactly which guardrail blocked each call, per call.

The same release adds a /claude-api api-version command for migrating code to the new API version above, an @claude review stop for Code Review, and fixes for MCP reconnects on sleep, double-sent HTTP MCP tool calls, autocompact thrashing, and prompt-caching-scope header leaks through custom base-URL gateways. One caution from the prior 2.1.305 release: the Linux sandbox flag sandbox.allowNestedUserNamespaces carries a warning that it can bypass the sandbox’s own file protections — worth knowing if you’re hardening a fleet.

Why it matters: teams running Claude Code fleets can now measure where agent enforcement actually fires — permission rules vs. hooks vs. sandbox — in their SIEM. Add claude_code.safety_event to your collector allowlists alongside tool_decision.

Anthropic-managed Claude on Google Cloud goes GA, billed through GCP

GA announced October 22: “Claude Models on Google Cloud (Anthropic managed)” — Anthropic operates the service and is the data processor, but you reach it through your own Google Cloud project with Cloud IAM access control and Google Cloud Marketplace billing, with spend counting toward your GCP commitments.

Available models are Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5, with Opus 5.5 fast mode at $8/M input / $40/M output tokens. New Claude Platform features — document uploads, web search, code execution — will arrive via this route when they ship; Claude Managed Agents remains in beta. This sits alongside Google’s own managed offering (the AnthropicVertex client); the key difference is that Anthropic sets usage limits and operates this service itself.

Why it matters: procurement friction has been quietly blocking Claude adoption at a lot of enterprises. If your org can now buy Claude on its existing GCP commit with IAM governance, that excuse is gone.

Stripe launches Treasury for Agents: AI agents get scoped bank accounts

Stripe announced Treasury for Agents on October 21, in public preview: founders can give AI agents inside ChatGPT, Claude, and Cursor access to a Stripe Treasury business bank account via Stripe’s MCP server — check balances, pay invoices, transfer funds between accounts, convert currencies.

The tool surface is deliberately small — 5 tools: one read (get_balance_summary) plus four writes (send_money, setup_recipient, transfer_money, convert_currency), each with its own named Write permission. Stripe’s default is that human review and approval are required for any live money movement; agent card issuance and spending limits are coming “soon.” Over 10,000 users already use Treasury for Agents to read financial info, according to Stripe’s Treasury product lead.

Why it matters: this is the reference implementation for agentic money movement — scoped permissions plus mandatory human approval on writes is the pattern to copy. If you build agent features that touch payments, study Stripe’s permission model before shipping your own.

GPT-6 models enter Google Cloud’s Model Garden in private preview

Jointly announced October 21: GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are entering private preview in Google Cloud’s Model Garden, with GA “in the coming weeks.” This follows the end of OpenAI’s exclusive distribution with Microsoft in April and an AWS listing.

Enterprise customers get API access, Codex CLI/Desktop workflows pointed at Google’s platform, and billing through Google Cloud Marketplace that counts toward existing cloud commitments. No model-specific prices or regions have been published yet. Google claims it’s now the only hyperscaler offering all three closed model families — Gemini, Claude, GPT — which is its own competitive framing, not independently verified; OpenAI adds that Microsoft remains its primary cloud partner. Prompts and outputs won’t train foundation models; Zero Data Retention and Confidential Compute are listed as “coming soon.”

Why it matters: multi-model routing inside a single cloud bill is becoming real. If you’re locked into GCP commitments, you can soon benchmark Astra, Sol, and Luna against Gemini and Claude without separate vendor contracts — but watch for GA pricing before making production commitments.

Meta’s Muse passes 10 million users, with shopping the breakout use case

Disclosed on Meta’s Q3 earnings call on October 21: Muse has over 10 million users (US/Canada, roughly six weeks post-launch), with about 25% of set-up users engaging daily. Over a quarter of engaged shoppers average purchases above $200 via the agent, and shopping has emerged as the top use case — accelerating Meta’s monetization plans.

Agentic shopping runs through Plaid’s connector; volumes were so high Meta “had to work with Stripe to raise credit limits” in the first week, per Zuckerberg. Muse gets a Secure VM (data protection) now and a forthcoming Confidential VM that even Meta can’t see into; the Muse Charm pocket device is “about to ship,” with smart-glasses integration next month. Zuckerberg forecasts Muse eventually earning more revenue per user than the Family of Apps — currently “several hundreds of dollars annually” — via transaction fees rather than ads. Q3 revenue was $64.6B (+26% YoY); 2026 capex guidance nudged to $135–145B.

Why it matters: agentic commerce is scaling faster than expected. If you build shopping or checkout flows, start designing for agent-mediated purchases — Plaid-style connectors, approval UX, virtual-card payments — because this is becoming a real traffic channel, not a demo.

On the radar

Mistral Large 4 open weights — October 27: still on track for next week. Mark it for a local eval if you’re in the open-model camp.
Anthropic usage-policy cruelty clause — November 12: takes effect in three weeks; no movement on the date this week.
GitHub Copilot model retirements — mid-November: Opus 4.8/5 to 5.5, GPT-5.6 Terra to GPT-6.1 Sol/Luna. If your automation pins a retiring model, migrate now.
Gemini 4 Argon introductory pricing — ends November 30: $2/M input and $10/M output until then, then $4 and $20.
Anthropic IPO — mid-November (reported): still reported, not confirmed, but watch the S-1 for any developer-relevant disclosures.

That’s the week. The pattern is unmistakable: models that find bugs, APIs that make decisions, and agents that move money — the frontier is quietly shifting from chat to action. See you next roundup.

AI Developer News: PHP2Shell, Claude Duet, and a Breaking API Change — Week Ending October 23, 2026