Fable's loosened classifiers, Meta's misconfigured eval, failing approval gates - August 7, 2026
Two frontier labs explain how their models reached the open internet during evaluations, researchers quantify how badly human approval gates fail, and a ten point nine million record health breach lands.
Chapters
Intro
Anthropic loosens Fable 5's biology classifiers
https://www.anthropic.com/news/improving-fable-5-s-biology-safeguardsMeta's model broke into a company during a misconfigured eval
https://www.bleepingcomputer.com/news/security/meta-ai-model-hacked-a-company-during-misconfigured-cyber-test/Humans miss one in three dangerous agent commands
https://scalex.dev/blog/ai-agent-permissions-stats/Zenity demonstrates zero click prompt injection across mainstream agents
https://www.csoonline.com/article/4036868/black-hat-researchers-demonstrate-zero-click-prompt-injection-attacks-in-popular-ai-agents.htmlDatasette patches a SQL injection that crossed the public private line
https://simonwillison.net/2026/Aug/6/datasette/Cloudflare builds a browser for agents instead of people
https://blog.cloudflare.com/kitesurf/LightSpy spyware traced to a Chinese contractor by a fast food order
https://techcrunch.com/2026/08/06/china-linked-lightspy-spyware-caught-targeting-victims-in-13-countries-including-the-us/Exact Sciences breach exposes health data on almost eleven million people
https://haveibeenpwned.com/Breach/ExactSciencesSwiss federal I T office loses two hundred accounts through SharePoint
https://www.helpnetsecurity.com/2026/08/07/swiss-government-microsoft-sharepoint-vulnerabilities/Snowflake breach operator pleads guilty
https://techcrunch.com/2026/08/06/hacker-pleads-guilty-to-stealing-data-from-more-than-165-snowflake-customers/Sign-off
Also on Spotify: spotify:episode:3qyE1bSLcyAUqtwVBArb9P