What Your Agent Remembers
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
This week the agent moved off the laptop, into hosted sessions, cloud sandboxes and a Git store built for thousands of agent pushes, while the week's trouble came from what agents remember for you: passwords saved by Claude Code's "don't ask again", stolen LLM keys running up six-figure days, and model strings that will quietly answer with a different model. Plus GPT-6 Cyber, Perplexity's open-weight embedders, Meta's Conductor shutdown and an MIT rebuild, Revolut bringing inference home, and the dates to act on.
Headlines
- Anthropic's amended S-1: internal estimates indicate it may imminently meet its RSP threshold for AI-accelerated progress TechCrunch
- METR's assessment: Anthropic's benchmark gains at about 1.9x the earlier rate, most likely mostly from AI, with much of the evidence redacted METR
- Binance's $6.43BN theft: a trojanized installer fetched while searching for an AI tool, a fake job interview, an abused internal AI tool The Hacker News
- ARTEX agents ran campaigns mostly on their own; one pulled 190,000+ game logins for an estimated $180–$360 of model spend Check Point Research
Anthropic
- "Don't ask again" can save the approved command, secret included; Apiiro found 1,122 public settings files holding at least one The Hacker News
- Claude Code v2.1.319: auto mode asks about or denies calls the server can't check when telemetry or feature flags are off Anthropic
- Message threads (beta): send only new messages; about 35% less wait on very large requests, Anthropic says The Next Web
- IP allowlist for the Claude Console (beta), and signing an Enterprise member out everywhere Anthropic
OpenAI
- GPT-6 Cyber: rated Critical for cybersecurity, 48.5% on ExploitGym, 81 Firefox flaws; Daybreak Red customers only BleepingComputer
- OpenAI's kernel run: 948 distinct root-exploitable flaws for about $1.2M in API credits; 61 patched upstream by the end of the run Help Net Security
- GPT-Image-2.6 Sunburst and Flare: adjustable complexity, SVG and animated GIF output OpenAI
- Outbound SIP calling in the Live API for all organizations OpenAI
- From December 1, Gemini 2.5 Pro and 2.5 Flash calls route to 3.8 Flash and 3.5 Flash-Lite, billed at their rates Google
- Gemini Embedding 2 Preview now routes to Gemini Embedding 2 (same schema, pricing and embedding space) Google
- Agent Substrate 1.0 and gVisor donated to the CNCF Google
- Antigravity v2.29.0: remote agents in the cloud, voice mode, Jupyter viewer Google
Grok / xAI
- Grok Imagine quality and pro image requests now served by grok-imagine-image-2.0-low xAI
- Musk: the cost of serving Grok Bot was unsustainably high relative to revenue The Next Web
Meta
- Meta buys Conductor; Conductor Cloud shuts down November 20 TechCrunch
- Conducted: an independent, MIT-licensed rebuild with a git worktree per task GitHub
Perplexity
- pplx-embed-v2-context: open-weight vision-language embedding models in 9B and 0.6B Perplexity
- The Decisions API can search the web and fetch pages before it answers ($0.001 per search) Perplexity
Elsewhere
- Mandiant: hijacked LLM access is generating fraudulent charges above $100,000 a day The Register
- Thinking Machines' Inkling 2 Small: 594B parameters, 17B active, Apache 2.0 weights promised after red-teaming Unite.AI
- Cognition's SWE-2 Flash, free in Devin through November 30 Unite.AI
- PyTorch 2.15's rewritten NCCL backend: one bad GPU need not restart the job The Register
Trend: the agent leaves the laptop
- Kimi Hosted Agents: managed sessions and cloud sandboxes for long-running agent work Unite.AI
- Cursor's cloud agent builds, checks and hosts a live app Cursor
- GitHub's new Git storage, built for thousands of agents pushing to one repository GitHub
- The other way: Revolut made a self-hosted open-weight model the default and cut AI coding inference costs about 90%, it says The Register
- Co-Design Bench: the best agent fully correct on 26.4% of real change requests The Decoder
Calendar
- November 16: Anthropic's second watermark wave (Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5) The Decoder
- November 16: NotebookLM Enterprise web addresses change; March 31, 2027: Gemini Code Assist entitlement ends Google
- December 3: Omni 1.1 Flash routes to Omni 1.2 Flash Google