Anthropic, Claude updates
New chatter in AI circles says Anthropic is keeping a powerful vulnerability‑finding model private under a project called Glasswing — a move that underscores how companies are treating potentially dangerous tools as internal-only for now — and Claude’s autonomous agent features have been moved onto paid plans, signaling commercialization of agent tooling. (x.com)
Anthropic did not just ship another chatbot feature this week. On April 7, 2026, it announced Project Glasswing, a program that gives a new unreleased model called Claude Mythos Preview to a closed group of defenders instead of the public. (anthropic.com) The reason is unusually blunt. Anthropic says this model can beat nearly all humans at finding and exploiting software vulnerabilities, and it says the model has already found thousands of high-severity flaws, including some in every major operating system and web browser. (anthropic.com) A software vulnerability is a hidden mistake in code, like a bad lock in a building that nobody noticed during construction. A model that can spot those locks at machine speed can help defenders patch systems faster, but the same skill can also help attackers break in faster. (anthropic.com) That is why Glasswing is structured like a restricted early-warning network, not a product launch. Anthropic says Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation are among the launch partners using the model for defensive work. (anthropic.com) Anthropic is also putting money behind the containment strategy. The company says it is committing up to $100 million in usage credits for Mythos Preview and $4 million in donations to open-source security groups, while extending access to more than 40 additional organizations that maintain critical software infrastructure. (anthropic.com) This did not come out of nowhere. On October 22, 2024, Anthropic introduced “computer use,” which let Claude look at a screen, move a cursor, click buttons, and type text, turning the model from a text generator into something closer to a remote operator. (anthropic.com) On May 22, 2025, Anthropic pushed further with Claude 4, saying Opus 4 and Sonnet 4 were built for long-running agent workflows, parallel tool use, and hours-long coding tasks. That matters here because a vulnerability hunter is not just answering questions; it is inspecting code, testing paths, and chaining many steps without constant human steering. (anthropic.com) At the same time, Anthropic has been moving those agent-style tools behind paid tiers. Its current pricing page lists Claude Code, Claude Cowork, and Research on Pro, while Max adds higher limits and early access to advanced features, and the Opus 4.6 page says the flagship model is available on Pro, Max, Team, and Enterprise plans. (claude.com) (anthropic.com) So the two updates fit together. The most dangerous agent capability is being kept inside a controlled consortium, while the safer commercial agent stack is being turned into a subscription business for individuals and companies. (anthropic.com) (claude.com) That is a pretty clear snapshot of where frontier artificial intelligence is in April 2026. The models are now good enough that one branch of the product line is sold like office software, while another branch is treated more like sensitive security infrastructure. (anthropic.com 1) (anthropic.com 2)
