Daily AI brief · September 14, 2026
Report: Contractors Read Real ChatGPT Conversations
404 Media reports that hundreds of OpenAI contractors read and grade real ChatGPT chats, with sensitive details sometimes getting past filters, so work data in personal accounts is now a live risk.
OpenAI's "Project Lily" puts real ChatGPT chats in front of contractors (404 Media) 404 Media obtained internal materials showing that hundreds of contractors, recruited through Crossing Hurdles and paid through Mercor, read real ChatGPT conversations and grade the model's responses. Reviewers don't see usernames. OpenAI tries to strip personal information first but acknowledged that sensitive details can still get through. What it means for SMBs / MSPs: Staff pasting client data into personal ChatGPT accounts is a data-handling exposure. Find out which plans your team actually uses and what each plan's terms say about human review.
Washington and Beijing both push back on "pacing" (Al Jazeera, NPR) President Trump posted that concern about AI is a "hoax" and rejected new guardrails. At a regular press briefing, China's Foreign Ministry spokesperson Guo Jiakun called Dario Amodei's essay fearmongering, pointing to its support for keeping chip export controls. AI-linked stocks fell on the day, according to Reuters. What it means for SMBs / MSPs: Don't expect a federal AI rulebook soon. Your compliance pressure will come from states, customers, contracts and insurers.
Agentic coding pushed Anthropic's CI load up 25x in six months (Claude blog) Anthropic says its engineers now ship about eight times as much code per quarter as they did from 2021 to 2025, and Claude writes roughly 80% of it. Test volume grew tenfold. Its first three patches to test selection lasted 70 days, 29 days and less than a day before it rebuilt the system. What it means for SMBs / MSPs: When a client adopts coding agents, the bottleneck moves to review, testing and build minutes, so budget for those before the invoice arrives.
ElevenLabs widens its hosted MCP into a creative connector (ElevenLabs) The hosted MCP server uses OAuth sign-in instead of pasted API keys. It now reaches 50+ models for voice, music, images and video from inside assistants like Claude and Cursor, and it can also create and edit voice agents in plain language. What it means for SMBs / MSPs: Marketing teams can produce voiceovers and audio from the assistant they already use. Set rules on brand voice and voice consent before anyone clones one.
ElevenLabs MCP inside Claude. It's for the owner or marketer who keeps putting off a product video or phone greeting because recording audio is a chore. A first test takes about 30 minutes. Connect the ElevenLabs connector with OAuth, paste in the copy from one existing product page, and ask for a 30-second voiceover in two stock voices. Listen for pronunciation of product and street names before anything goes public.
Ask each employee whether they use a personal or company AI account for client work. Reporting on OpenAI's Project Lily says hundreds of contractors read real ChatGPT conversations to grade responses. Usernames are hidden, but OpenAI acknowledged sensitive details can still get through. Move client work into a company-managed workspace, and add one line to your AI rules naming what never goes into any assistant.
*For MSPs:* An inventory of personal versus managed AI accounts, a check of each plan's terms and a one-page policy often leads to an account consolidation project.
Trusting "anonymized" as a synonym for "private." Removing a username doesn't remove the address, diagnosis or account number someone pasted into the chat. The reporting says sensitive details get past automated filters. If content would be a problem in a stranger's hands, it doesn't belong in a consumer account.
Data terms are becoming the buying criterion. Human review, retention windows, who holds the logs and whether you can prove deletion are moving from fine print to the first page of AI procurement. Vendors that give clear, plan-level answers, ideally with admin controls you can show an auditor, will win regulated and privacy-minded SMB buyers. The model leaderboard will matter less than those answers.
Compiled on September 27, 2026 from reporting published on or before September 14, 2026.