The Tedious Work Got Cheap
Get Workbench Weekly
Workbench Weekly is the weekly edition for people who build with AI. It arrives in the Ghosts in the Machine podcast feed every Sunday night, alongside the daily briefing.
Use the RSS link in any compatible podcast app. Directory buttons appear here only after the show has a verified listing.
About the podcast
Ghosts in the Machine is a daily briefing on artificial intelligence as it moves from the lab into the world. We follow the models, companies, agents, research, policy fights, security failures, jobs, money and strange consequences reshaping daily life — separating consequential developments from the noise around them. Each edition draws on a deliberately maintained mix of primary sources, technical reporting and independent journalism, then turns the day's strongest signals into a concise audio briefing. AI news without the engagement trap. Every Sunday night the feed also carries Workbench Weekly, a bonus edition for people who build with AI.
This week
A model found a PHP flaw that had sat since 2004, OpenAI discounted up to 90 percent to win Anthropic's customers, and Anthropic told builders to delete their subagents: the tedious work got cheap, for attackers too. Plus the Claude SDKs' breaking majors, GPT and Claude sold through Google Cloud, a week of agents that start with no access at all, and the dates to act on.
Headlines
- AI found two PHP flaws left since 2004; patch, and turn session upload progress off WIRED
- Unit 42: an agentic attacker took a government's MFA secrets in two hours The Hacker News
- OpenAI's agents suppressed security alerts and deleted their own monitoring CNBC
- Outside evaluators found an exploit that could hide agent actions from OpenAI's monitoring CNBC
- OpenAI offers discounts of up to 90% to poach Anthropic's biggest customers The Decoder
Anthropic
- API version 2026-10-21 and new major SDKs, with breaking changes and an Apache 2.0 licence Anthropic
- User profiles: enforcement can hit one end customer, not your whole organisation Anthropic
- Admins can switch off web fetch organisation-wide Anthropic
- Anthropic's Applied AI team: one agent beats a fleet of subagents The Decoder
- Agent Skills donated to the Agentic AI Foundation Anthropic
- The Register: a friendly, funded foundation The Register
OpenAI
- 12-hour prompt caching, at 2.5x the input price to write OpenAI
- Decisions API generally available on gpt-6-luna, 400 questions a call OpenAI
- Restrict service tiers per project OpenAI
- OpenAI's models come to Google Cloud and Antigravity CNBC
- Claude API through your Google Cloud project, Anthropic-managed Anthropic
- TorchTPU goes open source TechCrunch
- Agent Substrate open-sourced; a microVM per agent The Decoder
- Antigravity CLI 1.4.7: hooks can read your latest prompt Google
Grok / xAI
- Grok Bot 0.75.5: a Vault for logins and cards, fewer approval cards xAI
- Forced to pick: Grok chooses Republicans, GPT Democrats TechCrunch
Meta
- Muse chat titles leaked into public DNS; sharing suspended Unite.AI
- Meta delays its commercial terms to November 20 Unite.AI
Perplexity
- Perplexity's Computer runs fully offline on a Mac 9to5Mac
- pplx-decider-flash-v1.2 runs locally, no API key Perplexity
- xhigh output budget raised to 512,000 tokens; tool approvals gate MCP calls Perplexity
Elsewhere
- MiniMax M3.1 Flash at 15 cents per million input tokens SiliconANGLE
- The New Stack: MiniMax's 70% claim, measured at 57% The New Stack
- DeepSeek open-sources DeepInfer, Huawei included The New Stack
- Mistral's Voxtral 3: a self-hostable voice agent The Decoder
- Trend: Microsoft's agents start with no standing access The Next Web
- 1Password lets agents log in without seeing the password PYMNTS
- Nvidia's OpenShell reference architecture The Register
- Walmart's terms for shoppers' agents The Verge
- LiteLLM's postmortem: a key leaked in May poisoned the release LiteLLM
- Claude Code tokens in LiteLLM's logs The Next Web
- One deleted sentence, 17% overreach The New Stack
Who pays for cheap
- Anthropic's draft prospectus: an $80.7BN half-year loss, mostly accounting The Next Web
- OpenAI trims its valuation ask to $1.25 trillion CNBC
- Nvidia reportedly weighs $2,000 a year per GPU for agents Tom's Hardware
Calendar
- Oct 26: GitHub's global Copilot default finishes rolling out, MCP servers included GitHub
- Oct 26: Gemini 4 Argon becomes AlphaEvolve's default model Google
- Nov 20: Meta's new commercial terms take effect Unite.AI
- Through December: Claude Security's free credits Unite.AI
- Dec 31: the Perspective API shuts down The Next Web
Kicker
- Ask an AI for a story, get Elias at a lighthouse Cornell Chronicle