Disclosed critical CVEs hit 1,500 in June, 3.5 times the previous record
A historic spike in disclosed critical CVEs – over 1,500 in June alone, 3.5 times the previous monthly record – traced to AI-powered vulnerability discovery programs including Anthropic’s Project Glasswing, raising questions about responsible disclosure at scale. Mistral released Leanstral 1.5, an open-source formal math proof model that saturates the miniF2F benchmark and discovered previously unknown bugs in production code. AMD’s MI355X demonstrated 80% of NVIDIA Blackwell’s inference performance on GLM-5.2 at roughly half the cost, narrowing the gap between AMD and NVIDIA for production AI inference.
Security #
New Serious Vulnerabilities Spiked Around Release of Claude Mythos Preview #
Epoch AI / Hacker News (111 points)
Approximately 1,500 high- and critical-severity CVEs were publicly disclosed in June 2026 – more than 3.5 times the previous monthly record – following Anthropic’s announcement that Claude Mythos Preview could autonomously discover software vulnerabilities. Anthropic’s Project Glasswing, a partnership with Microsoft, Google, Apple, and AWS, has identified over 10,000 high- or critical-severity vulnerabilities, many still undisclosed, while OpenAI launched a similar initiative called Daybreak. The sheer volume challenges existing responsible disclosure processes and raises the question of whether the security ecosystem can absorb AI-discovered vulnerabilities faster than adversaries can exploit them.
Model Releases #
Leanstral 1.5: Proof Abundance for All #
Mistral AI / Hacker News (243 points)
Mistral released Leanstral 1.5, an open-source (Apache-2.0) formal math proof model for Lean 4 with 119B total parameters and 6B active. The model saturates miniF2F at 100% accuracy, solves 587 of 672 PutnamBench problems, and demonstrated practical value by discovering five previously unknown bugs in 57 open-source repositories – including an integer overflow in a cryptographic library. The three-stage training pipeline and strong test-time scaling from 44 to 587 solved problems as the token budget increases from 50K to 4M suggest that formal verification is becoming a tractable AI capability rather than an academic curiosity.
Infrastructure #
Performance Per Dollar Is Getting Faster and Cheaper #
Wafer AI / Hacker News (255 points)
Wafer demonstrated AMD’s MI355X GPU achieving approximately 80% of NVIDIA Blackwell B200 inference throughput on GLM-5.2 at more than 2x lower cost, reaching 2,626 tokens/second per node and 213 tokens/second single-stream. The team quantized GLM-5.2 from bf16 to MXFP4 using AMD Quark with lossless performance, and resolved two critical issues preventing speculative decoding that delivered close to a 3x single-stream throughput gain. The article argues that the AMD-NVIDIA gap is now “more a matter of support, not software,” meaning organizations with the engineering capacity to optimize AMD deployments can achieve substantial inference cost savings.
Probabilistic Memory Architecture for Trustworthy Edge Intelligence #
Semiconductor Engineering (Notre Dame, Georgia Tech, Villanova)
Researchers introduced p-MEM, a unified memory primitive that performs probabilistic sampling at native memory bandwidth, targeting Bayesian neural network inference on edge devices. The architecture reports reductions in instruction count, sampling latency, and energy for edge AI workloads. For teams deploying AI models on resource-constrained hardware where uncertainty quantification matters – safety-critical applications, sensor fusion, medical devices – in-memory probabilistic sampling could remove a bottleneck that currently forces approximate methods or cloud offloading.
Regulatory & Policy #
Fable #6: The Return of the King #
Don’t Worry About the Vase (Zvi Mowshowitz)
Zvi Mowshowitz analyzes the lift of US export controls on Claude Fable 5, which were imposed on June 12 after Amazon researchers discovered a potential vulnerability and restored by July 1 with stricter classifiers. The post criticizes the response as “entirely ad hoc,” argues the restrictions were counterproductive since debugging is fundamentally “fix this code,” and warns that OpenAI’s proposed 5% equity donation to a US sovereign wealth fund risks creating “ruinous” political capture of AI companies. The analysis captures the tension between legitimate security concerns about frontier model capabilities and a regulatory process that currently lacks the technical sophistication to calibrate its responses.
Developer Tools #
Guide to Running SOTA LLMs Locally #
jamesob / Hacker News (351 points)
A comprehensive guide for building high-performance local LLM inference rigs, with hardware specifications ranging from a $2K setup (Qwen3.6-27B on dual RTX 3090s) to a $40K system running GLM-5.2-Int8Mix at “almost-Opus” capability on four RTX 6000 Pro Blackwell cards with 384GB total VRAM. The guide emphasizes practical details: BIOS bifurcation settings, PCIe switch configuration for peer-to-peer GPU communication, kernel parameters for disabling IOMMU, and power limiting GPUs to 350W for circuit compatibility. For teams evaluating self-hosted inference versus API pricing, this provides concrete cost-performance reference points.
Fable’s Judgement #
Simon Willison’s Weblog
Simon Willison reports a key insight from his fireside chat with Cat Wu and Thariq Shihipar from the Claude Code team at AI Engineer World’s Fair: let Fable and Opus use their own judgment rather than dictating how they should work. The specific example was testing – Fable can decide when to use automated testing versus manual verification based on change complexity, often making better choices than prescriptive instructions. This aligns with the broader practitioner trend of treating frontier models less as instruction-followers and more as collaborators with genuine judgment about development methodology.
Understanding PyTorch’s Test Infrastructure #
PyTorch Blog
PyTorch published a guide explaining why test names in CI differ from source-level template names, covering the dynamic test generation system that expands device-generic templates across all device/dtype combinations at import time. The guide documents key decorators, OpInfo metadata, and the correct way to target specific tests with pytest filters rather than template names. For anyone contributing to or debugging CI failures in PyTorch, this demystifies a system that regularly confuses new contributors.
Open Source #
Open Source AI Gap Map #
Current AI / Simon Willison’s Weblog
Current AI, a non-profit founded at the AI Action Summit in Paris with $400 million committed, launched a Gap Map indexing the current state of open-source AI infrastructure and identifying where critical gaps remain. The project aims to build a “public option for AI” by systematically mapping what open-source alternatives exist for each layer of the AI stack and where proprietary solutions still dominate. For anyone making build-vs-buy decisions on AI infrastructure, the map provides a structured overview of where open-source options are mature versus where gaps remain.
Research & Papers #
AgRefactor: LLM Agents to Refactor Software for High Level Synthesis #
Semiconductor Engineering (Carnegie Mellon, UCLA)
Researchers introduced AgRefactor, an LLM-based multi-agent workflow for refactoring C/C++ software into HLS-compatible programs for FPGA and ASIC synthesis, reporting a 6.51x geometric mean speedup over a state-of-the-art pragma tuning tool. The system uses self-evolving agentic workflows to iteratively refactor code while preserving functional correctness. For hardware teams adopting high-level synthesis, this demonstrates that LLM agents can automate the labor-intensive refactoring step that currently bottlenecks the transition from software prototypes to hardware implementations.
Beyond Adam: SOAP and Muon for Faster Training of Machine Learning Interatomic Potentials #
Hugging Face Daily Papers
Researchers systematically compared matrix-structured optimizers (Muon, SOAP, and hybrid SOAP-Muon) against Adam for training NequIP and MACE interatomic potential models, finding that these newer optimizers deliver faster convergence and improved label efficiency. While the application domain is materials science simulation, the optimizer comparison is relevant to any team training transformer-based models where Adam is the unquestioned default. The results suggest that matrix-structured optimizers deserve evaluation beyond the domains where they were originally demonstrated.
Funding & Business #
AI Impact on Developer Education Sales #
Josh W. Comeau (Bluesky) / Simon Willison’s Weblog
Developer educator Josh W. Comeau reported that his new course launch is tracking to sell roughly one-third as many copies as a typical launch, with existing course sales also down significantly year-over-year. He attributes the decline primarily to AI: developers question whether skills are worth learning when AI can do the work, and those still learning increasingly rely on AI assistants rather than structured courses. This is a concrete data point on how AI is disrupting the developer education market – a leading indicator that may foreshadow broader impacts on professional training and certification businesses.
Other #
Google DeepMind and A24 Announce Research Partnership #
Google DeepMind
Google DeepMind and film studio A24 launched a multi-project research collaboration to integrate AI technology with creative filmmaking, with Google making a financial investment in A24. The partnership embeds A24 filmmakers directly with DeepMind researchers to co-develop tools for storytelling, positioning it as creator-centered AI development rather than technology imposed on artists. The collaboration signals growing investment in AI for entertainment production, though the practical impact on filmmaking workflows remains to be seen.
Threads to Watch #
AI-powered vulnerability discovery is creating a disclosure bottleneck. The 3.5x spike in critical CVEs from Project Glasswing and Daybreak represents a structural shift in software security: AI models that can autonomously discover vulnerabilities at scale generate disclosures faster than organizations can patch them. With over 10,000 vulnerabilities found but not yet publicly disclosed, the gap between discovery rate and remediation rate creates a window of risk that existing responsible disclosure frameworks were not designed to handle.
Open-weight inference economics are approaching a tipping point. AMD’s MI355X hitting 80% of Blackwell performance at half the cost, combined with a $40K local setup running “almost-Opus” quality inference, suggests that the hardware floor for competitive open-weight model serving is dropping faster than cloud API pricing. If self-hosting costs continue to compress while API margins remain high, the economic case for cloud-only inference weakens – particularly for organizations with the engineering capacity to optimize deployment.
Formal verification is becoming a practical AI capability. Leanstral 1.5 saturating miniF2F and solving 87% of PutnamBench problems while also finding real bugs in production code demonstrates that AI-assisted formal verification has moved from benchmark exercise to practical tool. Combined with AgRefactor’s agents for hardware synthesis, AI is expanding from code generation into domains where provable correctness – not just plausible correctness – is the standard.