Anthropic has introduced Claude Fable 5.1 and Claude Mythos 5.1, two models built on the same underlying technology but released under very different access rules. Fable 5.1 is aimed squarely at programming and complex knowledge work and is available the way prior Claude releases have been. Mythos 5.1, by contrast, is being made available only to a small set of verified organizations, reflecting its more advanced capabilities in cybersecurity and biological research.
The split marks a shift in how frontier labs are thinking about distribution. Rather than a single release cadence for every capability tier, Anthropic is now gating the most dual-use-capable version behind organizational vetting — treating deployment access, not just model training, as a safety lever in its own right.
The move lands as regulators and researchers increasingly press AI labs on how they handle models capable of meaningfully uplifting cyber-offense or bio-risk work, making the builder/restricted split as much a governance signal as a product one.
The Justice Department filed a brief in Manhattan federal court supporting OpenAI's argument that training large language models on copyrighted text generally qualifies as fair use. It is the first time the US government has weighed in on the wave of lawsuits brought by authors, publishers, music labels, and news outlets over AI training data — and it explicitly asks the court to reject any ruling that would treat LLM training itself as infringement.
Associate Attorney General Stanley Woodward tied the position directly to competitiveness, saying the administration will not let "our Nation be at a disadvantage relative to our foreign adversaries." Commerce Secretary Howard Lutnick separately urged G20 counterparts to embrace fair use for AI training while finding ways to compensate artists.
A government brief carries advisory rather than binding weight, and it doesn't resolve the Times' 2023 lawsuit outright. But it hands OpenAI and every other lab facing similar suits a powerful new argument, and signals where federal policy is likely to land as dozens of copyright cases against AI companies work through the courts.
Austin-based HiddenLayer closed a $100 million Series B led by Delta-v Capital, with Ten Eleven Ventures, Morgan Stanley, Microsoft's M12, and Booz Allen Hamilton also participating. The round pushes the four-year-old startup's total funding to roughly $156 million and follows annual recurring revenue that CEO Chris Sestito says grew more than 10x over the past year, with over 90% of that growth from new customers.
The company's platform scans AI file frameworks to catch models that aren't what they claim to be, protects deployed systems against prompt injection and model theft, and — with new funding — is extending runtime protection to autonomous coding agents specifically. One customer is described only as a "leading frontier model provider" with more than 700 million weekly users.
The raise arrives alongside a smaller but related deal: edge-security startup Huskeys closed a $27 million Series A from Blackstone to block agentic-AI attack traffic, underscoring how fast the AI-security funding category is filling out beneath the headline players.
Alibaba has released Qwen3.8-Flash-Next, a multimodal model built on a Mixture-of-Experts architecture that activates only part of the network for any given request rather than running the full model every time. The design is aimed at delivering competitive performance while sharply cutting the compute cost per query.
The release fits a broader pattern among Chinese labs of prioritizing cost-per-token and open access over raw parameter count, at a moment when US frontier labs are locking in tens of billions of dollars in compute contracts to train ever-larger models.
If efficient architectures like this one can hold their own on benchmarks, the practical effect is lower barriers to running capable AI — a dynamic that could matter as much for who gets to build on AI as which lab holds the top leaderboard spot.
Vara has announced CE certification for an AI system that can independently identify and triage mammograms read as normal — the majority of scans in any screening program, where radiologists currently spend most of their review time confirming nothing is wrong.
By handling the high-volume, low-risk portion of the workload autonomously, the system is designed to free up specialist capacity for the smaller share of cases that actually need close human attention, rather than simply flagging suspicious regions for a radiologist to double-check.
The certification is notable less for the specific use case than for the precedent: regulators approving AI to act, not just advise, inside a defined clinical pathway — a template other diagnostic-AI companies will be watching closely.