Today’s top line: A 2M‑prompt experiment found widespread prompt and credential leaks across AI routing services, while Meta says fast‑growing AI data‑centre emissions force a reassessment of its 2030 net‑zero target.
Researcher finds widespread secret leaks across AI supply chain
Summary: Researcher HD Moore and Dragos sent nearly two million prompts containing unique, trackable “canary” secrets to AI providers to test whether sensitive data leaks through AI routers and hosting services. The experiment found breaches of three hosting providers, with some stolen AWS keys reused “a couple hundred” times and the stolen bait keys used “tens of thousands” of times in AWS API calls.
Digest: The experiment shows that prompts and embedded credentials are exposed at multiple hops in modern inference reseller chains, creating real risk of theft and lateral access to cloud resources; businesses using AI routers should treat prompt data and embedded secrets as attack surface and audit all intermediaries.
Source: Ars Technica
Meta says soaring AI data‑centre emissions force 2030 rethink
Summary: In its 2026 sustainability report, Meta said demand for AI infrastructure has grown far faster than expected and it is re‑evaluating the 2030 timeframe for its net‑zero emissions goal. Meta reported total emissions rising from 15.6 million metric tonnes in 2024 to more than 21.7 million in 2025, a nearly 40 percent jump, and net emissions increasing from more than 8.1 million to more than 11.8 million metric tonnes.
Digest: Major cloud and AI compute expansion is forcing tradeoffs between speed and sustainability; firms that rely on hyperscale models should expect rising scrutiny on emissions and may need to adjust procurement, energy sourcing and reporting as timelines shift.
Source: Ars Technica
New method enforces physical bounds in AI models for chaotic systems
Summary: MIT researchers published a PNAS paper describing a way to add physics‑based restrictions to machine learning models so they never produce physically impossible, runaway predictions when simulating chaotic systems like weather or turbulent flows. The team says the method preserves the models’ ability to learn complex dynamics while keeping outputs within realistic bounds; code and data are being made available.
Digest: Reliable, bounded ML surrogates for chaotic systems could make AI more usable for weather, climate, aeronautics and industrial control, reducing risk from long‑run drift in deployed simulations.
Source: MIT News
MIT‑IBM team probes multimodal model failures, releases benchmarks
Summary: MIT News interviewed researchers from an MIT‑IBM collaboration that studied where state‑of‑the‑art multimodal foundation models fail to link perception (audio, vision) and language. The team developed new benchmarks and data‑efficient methods to recover latent multimodal correspondences without large annotated datasets, and has released those benchmarks publicly.
Digest: Work that exposes and benchmarks multimodal blind spots matters to businesses building audio/visual assistants and tools that must reason across modalities; expect improved evaluation suites and methods that reduce the need for costly labelled data.
Source: MIT News
Quick hits
- Google adds AI chat history sidebar, removes Search button – Google is rolling out a desktop redesign that replaces the old Search button with prominent AI feature buttons (image generation, file analysis, Gemini brainstorming) and an expandable left sidebar showing recent AI Mode threads on the homepage and results pages (Ars Technica).