The Weekly AI Digest
Week of 3–9 April 2026
Top 5 AI Stories This Week
Anthropic Withholds Claude Mythos, Launches Project Glasswing
Anthropic announced on April 7 that it will not publicly release Claude Mythos, the model leaked last week, due to its cybersecurity capabilities. Instead, the company launched Project Glasswing, giving a preview version to AWS, Apple, Microsoft, Google, Cisco, CrowdStrike, NVIDIA, JPMorgan Chase and the Linux Foundation — plus roughly 40 additional organisations that maintain critical software infrastructure. Anthropic is committing $100 million in usage credits and $4 million in donations to open-source security organisations. During internal testing, Mythos found thousands of zero-day vulnerabilities across every major operating system and web browser, including a 27-year-old bug in OpenBSD.
Why it matters: This is the first time in nearly seven years that a leading AI lab has withheld a model over safety concerns. For organisations running critical infrastructure, Project Glasswing is a signal to engage with AI-powered security testing now rather than waiting for attackers to get there first. For everyone else, basic cyber hygiene — patching, MFA, replacing unsupported devices — has never been more urgent.
Meta Releases Muse Spark — And It's Closed Source
Meta debuted Muse Spark on April 8, its first major AI model in over a year and the first product from Meta Superintelligence Labs, led by former Scale AI CEO Alexandr Wang. The model is natively multimodal, introduces a "Contemplating" reasoning mode that runs sub-agents in parallel, and achieves the same capability as Llama 4 Maverick using 10x less compute. Meta's stock rose roughly 10% over five trading days following the announcement. The catch: Muse Spark is entirely closed source, breaking with Meta's open-source Llama tradition.
Why it matters: For organisations that built workflows on open-source Llama models, Meta's shift to closed source changes the calculus on vendor dependency. For enterprises evaluating AI providers, Muse Spark narrows the gap with the leaders but doesn't close it — it trails on coding and abstract reasoning but leads on health-related tasks. The broader signal: the company spending $115–135 billion on AI capex this year has decided its best work is too valuable to give away.
OpenAI Closes $122 Billion Round at $852 Billion Valuation
OpenAI closed the largest private funding round in history on April 1, raising $122 billion at an $852 billion valuation. Amazon invested $50 billion (with $35 billion contingent on an IPO or achieving AGI), while NVIDIA and SoftBank each put in $30 billion. For the first time, OpenAI accepted $3 billion from retail investors through bank channels. The company says it now generates $2 billion in monthly revenue, serves 900 million weekly ChatGPT users, and is building a unified "superapp" combining ChatGPT, its Codex coding tool and a web browser.
Why it matters: For companies evaluating OpenAI as a vendor, the scale of this round means they are not going anywhere — but the IPO pressure also means product decisions will increasingly be driven by revenue metrics, not research ambitions. For competitors and startups building on AI, the concentration of capital is staggering: OpenAI alone raised more in this single round than most countries invest in technology in a decade.
Stanford AI Index 2026: China Closes the Gap, Public Trust Falls
Stanford HAI released its ninth annual AI Index on April 13, painting a picture of a field hitting breakthrough capabilities while raising urgent questions about transparency and public trust. Key findings: China has effectively closed the performance gap with the US on frontier models. On the SWE-bench coding benchmark, performance rose from 60% to near 100% in a single year. AI organisational adoption reached 88%. But model transparency is declining — average scores on the Foundation Model Transparency Index dropped from 58 to 40, with the most capable models disclosing the least. Only 10% of Americans said they were more excited than concerned about AI.
Why it matters: For organisations deploying AI, the transparency decline is a procurement risk — you are increasingly buying models where the developer won't tell you what data it was trained on or how large it is. For businesses worried about public perception, the Stanford data confirms that the gap between AI insiders and the general public is widening.
Wall Street Says the AI Bubble Already Burst — Quietly
Fortune reported on April 7 that the widely predicted AI stock market crash never arrived as a single event. Instead, Wall Street slowly and methodically wound down its euphoric AI investments over the better part of a year. NVIDIA's stock has stagnated for roughly three quarters despite continued earnings growth. Yet tech earnings remain strong: analysts project info tech EPS growth of 44% in Q1 2026, accounting for 87% of S&P 500 index earnings growth. Goldman estimates AI infrastructure investment will account for roughly 40% of all S&P 500 earnings growth this year.
Why it matters: For executives justifying AI investment to boards, the narrative has shifted from "get in early" to "show returns." The hype premium has deflated but the underlying economics remain strong. For businesses planning AI spending, this is arguably the healthiest possible environment: less froth, more focus on what actually works.
Australia Watch
Anthropic MOU Takes Effect as Amodei Wraps Canberra Visit
Following last week's signing, details of the Anthropic-Australia MOU continued to emerge. Anthropic will work with the AI Safety Institute on joint evaluations and share data on how AI is being used across the Australian economy. The company is exploring data centre and energy investments, and has allocated AUD$3 million in API credits to four Australian research institutions. Copyright law reform was a key discussion point during Amodei's meetings in Parliament House.
The partnership sets a template for how frontier AI labs may engage with middle-power governments on safety and economic development.
GovAI Chat Enters APS Trials
The Australian Government's GovAI Chat platform began trials across the Australian Public Service in April, providing government staff with AI tools that operate within Australian Government infrastructure and meet PROTECTED-level security requirements. The rollout is part of the APS AI Plan, which requires every agency to have a Chief AI Officer and mandates AI capability training for all public servants.
A significant step toward AI-enabled government services, with security and sovereignty built in from the start.
Emerging Trends
The "Too Dangerous to Release" Era Has Arrived
Anthropic withholding Mythos is a first for the current generation of AI labs. For businesses, the implication is twofold: the most capable AI models may no longer be available on the open market, and the cybersecurity threat from those that are is accelerating faster than most organisations' defensive capabilities.
Expect more capability restrictions as models approach and exceed human-level performance in sensitive domains.
Open Source Is No Longer the Default
Meta going closed-source with Muse Spark, after years of building its brand on open-weight Llama models, signals a broader shift. For businesses that built strategies around freely available frontier models, the assumption that the best models will always be accessible needs revisiting.
Vendor lock-in risk is increasing even for organisations that chose open-source specifically to avoid it.
The Trust Gap Is Becoming a Business Risk
Stanford's data shows only 10% of Americans are more excited than concerned about AI, while 84% of experts see it positively. For businesses rolling out AI-powered products or services to customers, this gap matters. The internal enthusiasm that drives AI adoption may not be shared by the people you are building for.
Customer communication about AI use is becoming as important as the AI itself.