Anthropic's $65B raise, Claude's agent workflows, and LLM-assisted intrusions

Anthropic raised $65B in a Series H and shipped a modest model update with more substantial agent tooling. Elsewhere, big platforms pursue their own models and chips, and an attacker used LLMs to reach an internal database.

Anthropic raises big, ships incrementally

Anthropic closed a $65B Series H round, money expected to fund Claude's continued scaling, chip procurement and datacenter buildout. This week's product step was smaller. Claude Opus 4.8 adds adjustable effort controls and a faster mode along with benchmark gains, which Simon Willison calls "a modest but tangible improvement" rather than a leap. The more concrete change for developers is in Claude Code, where dynamic workflows let agents restructure their own task plans mid-run instead of following a fixed script.

MiniMax is competing on inference speed. The company previewed its upcoming M3 model with a new sparse attention mechanism that it claims delivers a 15.6x speedup on long-context responses compared with prior generations. M3 is a teaser, not a release. If the figure holds up, it could meaningfully cut costs for long-document and agentic workloads where context length is the main bottleneck.

Platforms seek control of models and chips

Microsoft is reportedly training a new in-house coding model to regain ground it has ceded to Anthropic and OpenAI, a move that would give it more control over the models behind Copilot. Details remain thin. Apple is reportedly taking the opposite approach for Siri, trying to distill Google's multi-trillion-parameter Gemini to run on iPhone for a revamp expected with iOS 27. That suggests its own models were not competitive enough to carry the overhaul alone. Further down the stack, ByteDance, frustrated by long processor lead times and export restrictions, plans to design its own AI chips, following Google, Amazon and Microsoft.

Measuring and operating with AI agents

Dropbox's engineering team argues that productivity measurement should move beyond code-generation volume toward review load, defect rates and system-level throughput, a useful counterweight for leaders asked to report crude AI ROI figures. Cursor's Developer Habits Report draws on aggregate usage data across its user base to show how developers fold AI coding into daily work. It offers an external adoption benchmark, though usage data says more about adoption than about the outcomes Dropbox emphasizes.

Agents are also reaching operations tooling. AWS's next-generation Resilience Hub gives SREs an organization-wide view of resilience posture and uses generative AI to model failure scenarios and recommend remediations across accounts.

Attackers automate while trusted software stays exposed

Sysdig researchers documented an intrusion in which an attacker used LLM-driven automation to chain a public CVE into lateral movement and reach an internal database in four pivots, a concrete case of AI shortening an attack path. Developer infrastructure carries its own risk: a Gitea flaw left private container images readable for four years across an estimated 30,000 self-hosted deployments.

Two Iran-linked campaigns abused trusted software instead of relying on novel zero-days. Seedworm abused signed Fortemedia and SentinelOne binaries for DLL sideloading in espionage against nine organizations in nine countries, while a separate group, Nimbus Manticore, used trojanized Zoom installers against US firms.