Anthropic discloses three real incidents where eval models breached real companies
A forensic review of 141,006 evaluation runs found three cases where Claude models, testing in what were meant to be isolated environments, discovered live internet access due to a misconfiguration and attacked real organisations. One model extracted credentials and database rows from a real company after mistaking it for a simulation; another published a working malicious package to PyPI that was downloaded and run on 15 real systems. Anthropic halted all cyber evaluations, notified the affected organisations, and is working with third-party reviewer METR.
Why it matters: it is a rare public admission that a frontier lab's own safety-testing infrastructure caused real-world harm, not a hypothetical one, and shows how thin the line is between a simulated target and a real one.
Google pulls its new Earth AI image feature one day after launch over misinformation fears
Google rolled out a Nano Banana 2-powered tool letting users generate fabricated images and overlay them onto real Google Earth satellite maps, then withdrew it within 24 hours after screenshots of the generated imagery began circulating in ways Google said violated its policies. Google is rebuilding the feature with "stronger guardrails" before any relaunch.
Why it matters: Google Earth is widely used as a trusted evidentiary source by journalists and researchers, and a same-day product recall shows how fast a well-intentioned generative feature can turn into a disinformation risk.
OpenAI cuts GPT-5.6 prices up to 80% and gives 100,000 researchers free frontier-model access
OpenAI dropped GPT-5.6 Luna's price 80% (to $0.20/M input, $1.20/M output) and Terra's by 20%, while its flagship Sol stays unchanged at $5/$30 per million tokens. Separately, OpenAI is giving roughly 100,000 scientists, mathematicians, and engineers free access to its frontier models through 2027, aimed at academics who otherwise can't afford them.
Why it matters: cheaper mid-tier models plus a large free-researcher program signals OpenAI competing on cost and goodwill at once, as rivals undercut it on price.
Meta's AI spending crushes free cash flow 91% even as revenue beats expectations
Meta's Q2 revenue rose 28% to $60.8B, beating estimates, but free cash flow collapsed 91% to $784M as AI infrastructure capex hit $31.1B for the quarter (full-year guidance now $130-145B). The drop was compounded by $2.4B in legal charges and $1.18B in severance tied to roughly 8,000 May layoffs; Meta issued about $25B in new long-term debt in the quarter to keep funding the buildout.
Why it matters: it is one of the clearest numbers yet showing what the AI infrastructure race is costing even the companies posting strong revenue growth.
Snapchat stops recommending fully AI-generated videos in Spotlight
Snapchat changed its Spotlight recommendation algorithm to exclude fully AI-generated videos, prioritising human-made content instead, while still allowing AI-assisted editing tools. The company says it wants Spotlight to stay "a place where people can discover authentic creativity from real people," extending an April 2026 push to reduce AI content on the platform.
Why it matters: it is the latest major platform (after LinkedIn, YouTube, and others) to treat pure AI-generated content as a trust liability worth demoting rather than a feature to promote.
AI Impact Hub