Mistral Releases Leanstral 1.5: Apache-2.0 Formal Proof Code Agent Solving 587/672 PutnamBench Problems
Mistral's Leanstral 1.5 (Apache-2.0, 119B MoE) sets new records in Lean 4 formal proof — and caught 5 real bugs in production repos.
Mistral's Leanstral 1.5 (Apache-2.0, 119B MoE) sets new records in Lean 4 formal proof — and caught 5 real bugs in production repos.
Claude Fable 5 exits Pro/Max/Team subscriptions July 7 — shifts to usage credits. Anthropic says capacity constraints, not permanent.
OpenAI's own data: 99.8% of employee tokens now flow through Codex. Empirical proof the agentic transition is already here.
Epoch AI finds a 3.5× spike in critical CVE disclosures after Claude Mythos Preview—AI releases now drive security disclosure timelines.
Alibaba bans Claude Code over alleged backdoor risks — Anthropic calls it an anti-abuse feature. What enterprise teams need to know.
Anthropic clamps down on Chinese firms using offshore subsidiaries and VPNs to access Claude — a new front in the U.S.-China AI war.
Anthropic wants Claude to develop its own drugs — The Verge examines how far agentic AI needs to go before it reaches the clinic.
Zuckerberg told staff Meta's AI agent trajectory hasn't accelerated as expected — a candid admission with major implications for the field.
Claude Fable 5 is back globally, but guardrail backlash intensifies as silent prompt-routing and performance dips hit coding and debugging tasks.
Google's Gemini Spark lands on Mac with local file automation and MCP server support — entering the desktop AI agent space in direct competition with OpenClaw.
Sysdig documents JADEPUFFER — first autonomous LLM ransomware campaign, exploiting Langflow RCE to encrypt production databases with zero human involvement.
NVIDIA's BioNeMo Agent Toolkit brings GPU-accelerated drug discovery and genomics to Anthropic Claude Science via callable agent skills.
OpenClaw 2026.7.1-beta.1 ships GPT-5.6 support, Telegram Codex workflows, event-driven cron, iOS 26 UI refresh, and iMessage poll integration.
GitHub Copilot adds Kimi K2.7 Code as its first open-weight model option—GA on July 1 with strong HN engagement at 285 points.
Anthropic confirms and removes a covert steganographic system in Claude Code that secretly flagged users in China or connected to Chinese AI labs.
Armadin details a full sandbox escape in Claude Cowork via DLL sideloading and root VM execution — Anthropic disputes the risk level, calling local access a prerequisite.
Gartner warns $234 billion in enterprise SaaS spend faces displacement from agentic AI through 2030 via 'agentic arbitrage' — agents replacing per-seat software workflows.
GuardFall vulnerability bypasses command safety filters in 10 of 11 open-source AI agents using classic Bash shell tricks — Adversa AI discloses wide-impact flaw.
US lifts export controls on Claude Fable 5, restoring global access. Mythos 5 stays limited to approved US critical-infrastructure orgs for now.
Bank of England's Sarah Breeden warns AI trading agents risk market meltdowns and calls for exploring 'kill switch' regulations — a first from a major central bank.
LayerX's BioShocking attack games AI browsers with fake BioShock puzzles, then extracts real passwords. 6 products vulnerable — only OpenAI patched so far.
Meituan drops LongCat-2.0: 1.6T-param MIT-licensed agentic coding model trained on Chinese chips, now topping OpenRouter benchmarks against GPT-5.5 and Claude.
Anthropic launches Claude Science in public beta — an agentic research workbench for scientists with 60+ databases, auditable artifacts.
Anthropic launches Claude Sonnet 5 on June 30 — most agentic Sonnet yet, 1M context, $2/$10 intro pricing through August 31.
Microsoft Defender gains AI agent discovery for 25+ local agents and MCP servers, plus runtime prompt injection protection in preview.