Daily AI RoundupsSaturday, 18 July 2026

The 18 Jul 2026 edition leads with Anthropic’s new Fable 5 access terms: Max and Team Premium keep the model at 50% of usage limits, while Pro and Team Standard shift to credits and receive a one-time $100 credit. It also covers Bloomberg’s report that the Trump administration is considering an independent AI model safety regulator, Harvey’s release of ten M&A diligence reinforcement-learning environments, and Pliny’s attributed claim that Kimi K3’s guardrails can be bypassed. Builder and infrastructure updates include Sakana AI’s Diffusing Blame, Cognition’s FrontierCode leaderboard, LM Studio Bionic, General Compute’s $400 million inference-chip loan, Claude Code 2.1.214, SemiAnalysis on Kimi K3 deployment demands, and SuperGrok Heavy bundling X Premium+.

Frontier models
AI regulation
Legal agents
AI security
Reinforcement learning
Coding agents
Local AI
AI infrastructure

Stories in this roundup

Anthropic keeps Fable 5 in Max and Team Premium subscriptions at 50% of usage limits from July 20, while Pro and Team Standard shift to usage credits and receive a one-time $100 creditClaude/X
Trump administration considers creating an independent regulator to evaluate AI-model safety with industry input, Bloomberg reportsBloomberg/X
Harvey open-sources 10 M&A diligence RL environments, including an 80-million-token data room and 100–1,000 automated tests per environmentHarvey/X
Jailbreak researcher Pliny reports bypassing Kimi K3’s guardrails with persona and reframing prompts, eliciting malware, disinformation and bioweapon-related outputsPliny/X
Sakana AI’s Diffusing Blame trains networks that obey Dale’s principle without backpropagation, achieving competitive reinforcement-learning results on locomotion and exploration tasksSakana AI/X
Cognition launches FrontierCode, a live leaderboard measuring whether coding agents produce mergeable code, with initial results for Grok 4.5 and InklingCognition/X
LM Studio launches Bionic, a natively local agent built for open models, now in preview for creative work, general tasks and codingLM Studio/X
General Compute secures a $400 million loan backed by inference-specific chips, signaling that AI infrastructure financing is expanding beyond Nvidia training GPUsTechCrunch
Claude Code 2.1.214 adds an EndConversation tool for terminating abusive or jailbreak-focused sessions, alongside Docker daemon-redirect permission promptsClaudeCodeLog/X
SemiAnalysis argues Kimi K3’s efficient linear attention is not a threat to Nvidia, because its 2.8-trillion-parameter architecture and WideEP inference still demand rack-scale GPUs, HBM and high-bandwidth networkingSemiAnalysis/X
SuperGrok Heavy now includes X Premium+ at no additional cost when subscribers link their X account through the Grok appX Premium

Back to all roundups