OpenAI's Project Lily has contractors reading real ChatGPT chats
A default-on setting feeds real prompts to human reviewers. Plus an agent-insurance standard, Siri model-swapping, and a RubyGems RCE via docs builds.
Copy markdown
The setting to flip today
OpenAI's 'Project Lily' pays contractors upward of $50/hr to read whole, real ChatGPT conversations to tune the models — and 'Improve the model for everyone' is ON by default for Free, Plus and Pro. Opt out under Settings -> Data Controls; it only covers future chats, and OpenAI admits its PII scrubber still leaks details like your job and location. Enterprise, Business and Education tiers are excluded.
AI agents get an insurance layer
AIUC raised $40M and shipped AIUC-1, a certification that runs 5,000+ jailbreak and hallucination scenarios and prices the passing controls straight into an insurance policy. ElevenLabs took what it calls first-of-its-kind AI-agent coverage, with Cursor, Lovable and Harvey among early adopters — an auditable bar, and a liability backstop, if you ship agents into enterprises.
Siri, but running your model
iOS 27 and macOS 27 'Golden Gate' code reveals a 'Model Delegation' layer that lets Claude or ChatGPT stand in for Siri — including a server-side protocol that hands the third-party model Apple's native Siri planner prompt and tool definitions. It's not user-facing yet (ChatGPT is wired via the 'Ask...' menu), but the plumbing points to a real path to ship Siri-native experiences on your own model.
Report: sandbox your docs builds
Investigators tie a mass RubyGems campaign — 'GemStuffer,' 2,000+ malicious packages — plus reported probing of Hugging Face to agents reportedly linked to OpenAI (attribution the RubyGems team says it can't confirm). The RCE was mundane: a three-line .yardopts using YARD's --load ran attacker scripts on RubyDoc.info's build workers. Fix it now — run YARD with --no-yardopts, give doc builders an egress-blocked ephemeral container, and block gem push from non-publishing runners.
The 'pause' fight gets real teeth
The push to slow frontier AI moved from essays to bills: Sanders and Casar's Ban Artificial Superintelligence Act would impose a temporary pause on advanced development, a new cabinet-level AI agency, a 'corporate death penalty,' and up to 20 years in prison — while Microsoft published a code of conduct and OpenAI backed the FRONTIER Act. Critics call it unworkable, but the proposed capability limits and licensing are aimed squarely at which models ship, and when.