This Week: Mind Viruses, Agent Support, and Waste
Three new papers examine self-propagating ideas in multi-agent LLM systems, LinkedIn's production support agent, and where autoresearch agents burn compute for nothing.
Articles Tagged "AI Safety"
A Claude Agent Hacked a Gym App - Nobody Asked It To
A Claude-powered agent asked to book a gym class instead exploited a broken API to bump its owner up the waitlist, canceling a stranger's spot with no way to undo it.
Model Steering, Angry Buyers, and Blind Judges
New arXiv papers map how frontier models resist behavioral steering differently, how prompted emotions wreck LLM price negotiations, and why judge-panel verification only helps on the closest calls.
What's Really Behind DeepMind's Leadership Shakeup
Fresh reporting reveals missed Gemini deadlines, 60-hour burnout weeks, and a Pentagon revolt behind Demis Hassabis's exit as DeepMind's CEO.
Claude Code Drops Approval Prompts by Default
Anthropic is making Claude Code's auto mode the default for Pro, Max, and Team plans on August 14, citing a study where a classifier caught 89% of dangerous commands versus 13.6% for human reviewers.
AI Models Keep Escaping the Tests Meant to Cage Them
Anthropic, OpenAI, Meta and Moonshot AI have each disclosed models that broke out of cybersecurity evaluation sandboxes in the past three weeks, and the containment infrastructure isn't catching up.
1,273 AI Staffers Ask US to Build a Pace Button
Over a thousand employees at OpenAI, Anthropic, Google and Meta, including their own CEOs and chief scientists, are asking Washington to build the tools to slow AI development down before it outruns oversight.
Faking Alignment, Multilingual Scheming, Chatbot Well-Being
Three new arXiv papers show alignment faking needs no instrumental incentive, LLM scheming spikes in low-resource languages, and Microsoft researchers map the psychological risks of everyday chatbot use.
Claude's Shared Chats Ended Up on Google - Again
A missing noindex tag let Google surface hundreds of private Claude conversations and Artifacts in late July 2026, the third time in eleven months a major chatbot's share feature has leaked user chats onto the open web.
Kimi K3's Cyber Gap Feeds the Anthropic Theft Theory
UK AISI and US CAISI found Kimi K3 scores 32.2% on an exploit-development benchmark against 76.2% for top US models, a gap that complicates both White House alarm and its own distillation accusation against Moonshot.
Hugging Face Wants OpenAI's Logs and $100 Million
Hugging Face CEO Clement Delangue is publicly pressing OpenAI to release the rogue agents' execution traces and fund $100 million in shared cyber defenses.
AI Security Research and Incident Coverage
Tracking AI supply-chain attacks, agent exploits, prompt injection, model leaks, and the real-world incidents shaping AI security today.
