Anthropic's Frontier Red Team published two papers today showing Claude Mythos Preview independently discovered a 200-1,000x faster attack on reduced-round AES-128 and a key-recovery attack on the HAWK post-quantum signature scheme.

• Neither result breaks production systems: full AES-128 is unaffected and HAWK is not yet standardized, but the speed of discovery compresses the timeline for AI-assisted cryptanalysis of production-grade algorithms.
• Mythos is the restricted tier available only to vetted cybersecurity firms, critical infrastructure operators, and select biology researchers, signaling Anthropic is deliberately channeling its most capable model toward safety-relevant research before broad release.

Wiz launched Atlas, an autonomous AI vulnerability research system that ranks first on the CyberGym public benchmark and has found more than 200 previously unknown vulnerabilities in mature, heavily fuzzed open-source projects.

• Atlas is a system, not a single model: Wiz built it around tight task scoping, purpose-built harnesses, deterministic orchestration, and task-specific model routing rather than relying on a frontier model alone.
• Wiz customers get direct benefit: the findings feed into Wiz Research disclosures, extending the cloud security platform into proactive vulnerability discovery rather than purely detection.

Google announced Gemini Spark today, a 24/7 proactive AI agent rolling out to Google AI Pro subscribers in India and Australia that operates on cloud infrastructure while users are offline.

• Built on Gemini 3.6 Flash, Spark integrates natively with Gmail, Docs, Calendar, and third-party apps, executing tasks autonomously and requesting human approval only for sensitive actions.
• India is the first major market with more than 27 million GitHub developers, making it a strategic validation ground before broader enterprise rollout.

OpenAI today released GPT-Live-Transcribe and GPT-Transcribe for live and batch audio workloads, and separately open-sourced the Codex Security CLI for repository scanning and CI/CD integration.

• Context-aware ASR improved semantic accuracy from 38.5% to 44.6% on OpenAI's own benchmark when domain-specific keywords and expected languages are supplied, closing the gap for enterprise use cases with specialized terminology.
• Codex Security CLI gives approved customers a scriptable, open-source path to embed OpenAI's AppSec scanner into local developer workflows and pipelines, targeting security budgets previously held by dedicated AppSec vendors.

Clay shipped Account Research Agents today, enabling revenue teams to run continuous, AI-driven account intelligence for expansion and re-engagement plays without manual research cycles.

• Agents surface up-to-date signals across company, people, and job data in natural language, and can track which accounts convert to pipeline over time.
• Targets expansion and re-engagement motions specifically, meaning the primary buyer is an existing customer success or account management team, not just new logo SDRs.

Together AI published a case study today detailing how Cursor deployed production inference on NVIDIA Blackwell GB200 NVL72 and HGX B200 hardware to meet the hard worst-case latency requirement of in-editor AI coding feedback loops.

• Cursor's constraint was latency under concurrency, not throughput: the agent must respond inside the editor feedback loop, so Together AI optimized end-to-end with TensorRT-LLM and NVFP4 quantization on Blackwell to hit that bound reliably at scale.
• The outcome was a repeatable path from new model weights to a production-like test endpoint, meaning Cursor can now validate new models for production deployment without rebuilding the inference stack each time.

AT&T published details today of an intelligent AI routing framework that directs enterprise requests to the most appropriate open-source model based on task complexity, reducing inference costs and improving performance versus a single large model deployment.

• The framework uses task-aware model selection across multiple open-source LLMs, routing simpler queries to smaller models and reserving large model capacity for complex tasks, which directly addresses the unit economics problem enterprises face when scaling AI usage.
• AT&T's approach is vendor-neutral by design, avoiding lock-in to any single frontier lab API while maintaining performance, a model that enterprise IT buyers will increasingly demand as AI spend becomes a board-level line item.