Mythos Finds Zero-Days in Every Major OS and Browser

Anthropic announced Project Glasswing, a coalition with AWS, Apple, Google, Microsoft, Nvidia and seven other organisations using Claude Mythos Preview to find and fix vulnerabilities in critical infrastructure. The model has already identified thousands of high-severity flaws in every major OS and browser. Anthropic is committing $100M in usage credits and $4M in direct donations to open-source security groups. Over 40 additional organisations have received access.

OpenAI released "Industrial Policy for the Intelligence Age," proposing a Public Wealth Fund giving every citizen a stake in AI-driven growth, a shift from payroll to capital gains taxation, and 32-hour four-day workweek pilots. The document also calls for treating AI access as a right, portable benefits decoupled from employers, and model-containment playbooks for dangerous systems. OpenAI is backing it with $100K research grants and up to $1M in API credits.

Intel is channelling billions into advanced chip packaging at facilities in New Mexico and Malaysia, positioning the once-dormant Fab 9 as critical AI infrastructure. CFO Dave Zinsner says packaging revenue projections have shifted from "hundreds of millions" to "well north of $1 billion," with deals in the billions per year close to closing. Multiple sources say Google and Amazon are in ongoing talks as customers.

Nvidia's acquisition of SchedMD, the company behind the Slurm workload manager, is drawing concern from AI industry executives. Slurm runs on roughly 60% of the world's supercomputers and is used by Meta, Mistral, and Anthropic for model training. Analysts warn Nvidia could subtly favour its own hardware through roadmap decisions, noting CUDA ecosystem support already ships faster than AMD's ROCm or Intel's oneAPI alternatives.

Axios reports Meta is preparing to release the first AI models developed under Alexandr Wang, with plans to eventually open-source versions. The company is shifting to a hybrid strategy where some of its largest models stay proprietary while others ship openly. Wang sees Anthropic and OpenAI as focused on enterprise and government, while Meta targets consumers through WhatsApp, Facebook, and Instagram integration.

Anthropic's red team benchmarked Mythos Preview against Firefox 147 and found it generated 181 working exploits compared to Opus 4.6's two. The model found a 27-year-old OpenBSD bug and a 16-year-old FFmpeg vulnerability that decades of fuzzing missed. It autonomously built 20-gadget ROP chains and escaped browser sandboxes through JIT heap sprays. Manual validation showed 89% severity agreement with human experts.

Caer Sanders distils mechanical sympathy into four everyday principles: predictable memory access, cache line awareness, the single-writer principle, and natural batching. The concept, borrowed from Formula 1 and popularised in software by Martin Thompson, explains why the LMAX architecture processes millions of events per second on a single Java thread. Sanders applies these patterns to AI inference platforms and distributed data systems.

Al Chen, a field engineer at Galileo, uses Claude Code to query the company's entire codebase across 15 repositories, combined with Confluence and Slack via MCP. A 16-line script pulls the latest main branch from all repos every morning, ensuring answers come from current code rather than stale docs. He maintains a "customer quirks" page so Claude Code generates deployment instructions tailored to each enterprise client's specific setup.

Traditional uptime monitoring fails for AI agents because they don't crash when they break. They hallucinate policy details, lose conversation context, or burn through token budgets while dashboards show 100% availability. This guide distinguishes "system uptime" (endpoints return 200s) from "functional uptime" (accurate, cost-effective outputs) and proposes a three-tier approach covering infrastructure, orchestration continuity, and agent-level behavioural monitoring.

An engineer needed to extract revision numbers from 4,700 engineering drawing PDFs. The manual alternative was 160 person-hours at £50/hour. Rather than throwing everything at GPT-4 Vision, the solution used a hybrid pipeline: deterministic text extraction for the 70-80% of documents with parseable text, and LLM vision only for scanned legacy drawings where Python regex couldn't reach. The system design mattered more than the model choice.

In a multi-agent fraud detection pipeline, switching from REST/JSON to gRPC bidirectional streaming cut orchestration latency from 2.1 seconds to 420ms. The "JSON tax" compounds across agent hops: serialisation overhead, payload bloat (protobuf encodes the same data in 60-80% fewer bytes), synchronous lock-step communication, and zero schema validation. For pipelines where five agents need to coordinate in real time, the wire protocol matters more than the model.

Metronome CEO Scott Woody argues AI agents are driving a "third era" of software pricing, moving from perpetual licences to SaaS subscriptions to consumption-based billing. When thousands of agents run in the background burning API tokens, per-seat pricing stops making sense. Woody, who previously led monetisation engineering at Dropbox, says customers demand real-time visibility into spend, predictable costs, and usage controls before they'll trust consumption models at scale.

Helen Toner writes that the term AGI has become a "fuzzy cloud" that obscures more than it clarifies. Serious people simultaneously claim AGI arrived in 2025 and that it's a decade away, because different definitions (matching human performance, exceeding it, economic value, learning efficiency) point to wildly different timelines. With AI capabilities now jagged rather than uniform, the single-milestone framing has broken down.

GitHub saw a 14x annualised increase in commits over three months, largely driven by coding agents, and the supply side can't keep up. Nvidia's liquid-cooled GB200 rollout has been painful at gigawatt scale, DRAM fabrication bottlenecks are tightening memory supply, and signing $100B compute commitments does not magically create capacity. Martin Alderson's analysis suggests the crunch is structural, lasting 18 to 24 months.

Hippo is a cross-tool memory layer for AI coding agents, built on SQLite with markdown/YAML mirrors that stay git-trackable. It imports context from ChatGPT, Claude, and Cursor so switching tools doesn't mean starting from zero. The system uses reward-proportional decay inspired by spiking neural networks: memories with positive outcomes decay slower, negative ones faster. Public benchmarks show 74% retrieval accuracy on the LongMemEval standard.