Formal verification has long been the territory of academic mathematicians and safety-critical embedded systems engineers — deeply important work, but rarely the domain of general-purpose AI. Mistral is changing that narrative with Leanstral 1.5, a 119B Mixture-of-Experts model released under Apache-2.0 that doesn’t just write code in Lean 4 — it operates as a genuine agentic system, editing files and running bash commands to verify proofs. The benchmarks are remarkable. Leanstral 1.5 solves 587 out of 672 PutnamBench problems (a benchmark of undergraduate competition mathematics problems), achieves an 87% score on FATE-H, and saturates miniF2F — a standard Lean formalization benchmark. More practically, it was deployed against 57 real-world repositories and found 5 previously unknown bugs. ...

The AI scaling story just got more complicated — in the best possible way. Shanghai AI Lab (InternScience) has open-sourced Agents-A1, a 35-billion-parameter Mixture-of-Experts model that matches or beats trillion-parameter systems on agentic benchmarks. The kicker? It does this not by throwing more parameters at the problem, but by doing something far more interesting: training on dramatically longer agentic task trajectories. The paper is titled “Scaling the Horizon, Not the Parameters” (arXiv:2606.30616). That subtitle is the entire thesis. ...

Claude Fable 5 Is Leaving Subscriptions July 7 — What You Need to Know Claude Fable 5 just returned from a 19-day export control pause on July 1. Now, three days after its restoration, Anthropic has announced that it will leave Claude subscription plans on July 7 — less than a week after coming back. This is not the same situation as the export control removal. That was a regulatory hold. This is a capacity decision. And according to Anthropic, it’s temporary. ...

Claude Fable 5’s Safety Classifiers: A Field Guide to the Opus 4.8 Fallback After 19 days of unavailability due to export control review, Claude Fable 5 was redeployed on July 1, 2026, with a significant architectural addition: upgraded safety classifiers that detect when an agent loop is handling cybersecurity or biological topics and route those tasks to Claude Opus 4.8 instead. For most general-purpose applications, you’ll never encounter this. For operators building security agents, penetration testing assistants, threat intelligence tools, or any application touching biological research — it’s a core operational reality you need to understand. ...

OpenAI’s Landmark Study: 99.8% of Employees Now Route Work Through Agentic AI We’ve all heard the “agentic AI is the future” prediction so many times it’s become background noise. OpenAI’s new research paper — “The Shift to Agentic AI: Evidence from Codex” — is different. It doesn’t make predictions. It presents data. Real usage numbers from real workers across three groups: individual users, organizational users, and OpenAI’s own 3,000+ employees. ...

Protecting Your Agent Sessions: Claude Code Workspace Isolation Best Practices On July 4, a GitHub issue filed against the Claude Code repository — issue #74066, “[Bug] Potential session/cache leakage between workspace instances or consumer accounts” — trended to #1 on Hacker News with 54+ points within its first hour. As of this writing, Anthropic has not confirmed the bug. But here’s what’s not in dispute: Claude Code has had prior documented session isolation issues. GitHub issue #29342 documented cross-session transcript leakage where entries were written to the wrong JSONL file. That issue was resolved, but the pattern of cross-session contamination is now a twice-appearing vulnerability class. ...

ZCode for Enterprise: Data Sovereignty Checklist Before Switching from Claude Code or Cursor Z.ai just launched ZCode as a genuinely compelling free alternative to Cursor and Claude Code. It’s powered by GLM-5.2’s 1M context window, ships on macOS, Windows, and Linux, includes a multi-agent “Goal Mode,” mobile bot control, and a plugin architecture — all at zero cost for the daily free quota tier. For individual developers, it’s hard to argue with the price. ...

When Claude Fable 5 went back online on July 1, 2026, it came with something the AI security community had never seen before: a formally documented, standardized framework for rating the severity of AI jailbreaks. Anthropic published its Cyber Jailbreak Severity (CJS) Framework alongside a HackerOne bug bounty program specifically for cybersecurity jailbreak reports. If you’re building red-team pipelines, security evaluation workflows, or responsible disclosure processes around frontier AI, this framework is the new baseline. ...

If you’re running automated pipelines with Claude Code, the July 3, 2026 release of v2.1.200 includes two changes that will affect how those pipelines behave — and one of them could silently break workflows you haven’t tested recently. Both changes are safety improvements that make Claude Code more predictable and transparent in automated contexts. But they require action if you’ve been relying on the old behavior. Here’s what changed and what to do about it. ...

Something remarkable happened to the global software security landscape in June 2026. As Anthropic’s Claude Mythos Preview began reaching its initial wave of defensive partners, organizations started racing to find and disclose their vulnerabilities — not because regulators demanded it, but because frontier AI was about to make exploitation trivially easier. Epoch AI has now documented this phenomenon in striking detail: a 3.5× surge in high-severity and critical CVE disclosures in a single month, the largest on record. ...