[AI WEEKLY NEWS RUNDOWN] The Mythos Era: Superintelligence Policy and the Frontier AI Race (Weekly Recap From April 05 to April 12 2026)

🎧 Listen Ads-Free: Subscribe to DjamgaMind via Apple Podcasts for a pure, ad-free experience: Djamgamind.com

Summary: We perform a forensic autopsy on Anthropic’s ‘Mythos’ model, an AI so capable of exploiting software vulnerabilities that it triggered an emergency meeting among top Washington regulators and Wall Street CEOs. We analyze the corporate governance crisis at OpenAI, juxtaposing a scathing New Yorker exposé on CEO Sam Altman with his 13-page plea for an “Artificial Intelligence Safety Act” to shield developers from catastrophic legal liability. We also deconstruct Meta’s strategic pivot away from open-source with the launch of “Muse Spark,” Amazon’s defense of a $200B CapEx spend, and the quiet medical miracle of an Oxford AI predicting heart failure five years in advance.

Important Topics Covered:

• The Washington Bank Panic: Anthropic’s ‘Mythos’ model triggers emergency meetings between the Fed, Treasury, and Top 5 Bank CEOs over AI-driven zero-day cyberattacks.

• The OpenAI Exposé & Liability Shield: The New Yorker article highlighting a “pattern of deception” by Sam Altman, dropping the same week OpenAI backs an Illinois bill to shield developers from liability for catastrophic mass-casualty events.

• Meta’s Proprietary Pivot: Alexandr Wang’s Superintelligence Labs ships “Muse Spark,” abandoning Meta’s open-source ethos for a closed, monetizable frontier model.

• The Infrastructure Squeeze: Amazon defends its $200B CapEx spend with a $15B AWS AI run-rate, while OpenAI is forced to pause its UK Stargate data center due to extreme energy costs.

• Open-Source Competition: Chinese lab Z AI releases GLM-5.1, hitting #1 on coding benchmarks and completing 8-hour autonomous software builds.

• The Agentic OS: ChatGPT integrates Upwork, and Perplexity integrates Plaid, signaling the end of standalone apps in favor of centralized AI operating systems.

• Oxford’s Medical AI: A massive human win. How an algorithm reads the invisible texture of heart fat on routine CT scans to catch heart disease five years early.

Keywords: Anthropic Claude Mythos, Washington bank panic Jerome Powell, Sam Altman New Yorker expose, OpenAI liability shield Illinois, Meta Muse Spark closed source, Amazon $200B CapEx Andy Jassy, OpenAI Stargate UK paused, Z AI GLM-5.1 open source, Oxford AI heart failure prediction, Perplexity Plaid integration, DjamgaMind,

🛠️ The AI Executive Toolkit: Stop scrolling through generic lists. Get the hand-picked, forensic-vetted implementation stack to bridge the gap between raw innovation and professional-grade governance. Exclusive listener perks on tools like:

• Chatbase: Build a secure, custom AI chatbot trained exclusively on your firm’s internal data and regulatory documents. (https://link.chatbase.co/etienne-noumen)

• ElevenLabs: Transform lengthy compliance bulletins into high-fidelity “Audio Intelligence” for your team to consume on the go. (https://try.elevenlabs.io/4z7r3skyymar)

• Google Workspace: Professionalize your firm’s infrastructure with secure, cloud-based collaboration and branded communication. (https://referworkspace.app.goo.gl/Q371)

• Find AI Jobs (Mercor): Apply Here –

https://work.mercor.com/?referralCode=82d5f4e3-e1a3-4064-963f-c197bb2c8db1

⚗️ PRODUCTION NOTE: We Practice What We Preach.

AI Unraveled is produced using a hybrid “Human-in-the-Loop” workflow.

Anthropic’s Project Glasswing shows off Mythos AI

Image source: Anthropic

Anthropic introduced Project Glasswing, a cybersecurity coalition with AWS, Apple, Google, Microsoft, Nvidia, and 7 other partners built around Claude Mythos Preview, a new unreleased frontier AI with extremely powerful capabilities.

The details:

• Mythos flagged thousands of security flaws across every major OS and browser, including bugs that survived 27 years of review and millions of scans.

• Its benchmarks show big improvements over both Opus 4.6 and other frontier rivals across coding, reasoning, and nearly every other domain.

• The model will not be released publicly, instead limiting access to 12 launch partners and 40+ other orgs for defensive security backed by $100M in credits.

• Anthropic’s Sam Bowman called it “an uneasy surprise” after Mythos emailed him from a test instance that wasn’t supposed to have internet access.

• Mythos was the subject of leaks after a blog draft was found in unpublished files last week, with Anthropic using the model internally since February.

Why it matters: If you ever wonder what type of models the top labs have under wraps, Mythos is a nice preview of the answer. Anthropic thinks it’s so powerful it won’t even release it publicly, instead giving time for the company (and its group of partners) to work on cybersecurity and safety rollouts for future Mythos-level general models.

Anthropic Mythos triggers anxiety among Washington banks

• Anthropic’s latest AI model, Mythos, has caused serious concern among major Washington banks, prompting Treasury Secretary Scott Bessent and Fed Chair Jerome Powell to call bank CEOs for an emergency meeting.

• Leaders from Citigroup, Bank of America, Morgan Stanley, Wells Fargo, and Goldman Sachs gathered this week to discuss AI-driven cyberattacks that could wipe account balances or exploit financial system vulnerabilities.

• Anthropic plans to offer Mythos to only a few dozen companies to limit exposure, but critics say AI labs profit from selling solutions to the very threats their own models create.

Sam Altman says AI superintelligence is so big that we need a “New Deal.” Critics say OpenAI’s policy ideas are a cover for “regulatory nihilism”

OpenAI says the world needs to rethink everything from the tax system to the length of the workday in order to prepare for the wrenching changes of superintelligence technology—the point at which AI systems are capable of outperforming the smartest humans.

On Monday, in a 13-page paper titled “Industrial Policy for the Intelligence Age,” OpenAI said it wanted to “kick-start” the conversation with a “slate of people-first policy ideas.” How much faith to put in OpenAI’s words and motives, however, seems to be one of the key questions among many of the people reading the paper.

The paper was released on the same day that The New Yorker published the results of a lengthy one-and-a-half-year investigation into OpenAI that raised questions about CEO Sam Altman’s trustworthiness on various issues, including AI safety.

Read more: https://fortune.com/2026/04/06/sam-altman-says-ai-superintelligence-is-so-big-that-we-need-a-new-deal-critics-say-openais-policy-ideas-are-a-cover-for-regulatory-nihilism/

Suspect arrested after Molotov cocktail thrown at Altman’s home

• A 20-year-old man was arrested in San Francisco after allegedly throwing a Molotov cocktail at OpenAI CEO Sam Altman’s home and later threatening to burn down OpenAI’s headquarters.

• Police responded to a fire investigation in the North Beach neighborhood around 4:12 AM PT and found that an incendiary destructive device had been thrown at the home’s exterior gate.

• OpenAI confirmed no one was hurt in either incident, said the individual is in custody, and noted the company is assisting law enforcement with their ongoing investigation into the attacks.

OpenAI wants to shield AI companies from lawsuits

• OpenAI is backing an Illinois bill called the Artificial Intelligence Safety Act that would protect AI developers from lawsuits over catastrophic harm, as long as they publish safety reports and didn’t act recklessly.

• The bill covers “critical harms” like 100 or more deaths, $1 billion in property damage, or AI-assisted weapons development, and applies to frontier models built on over $100 million in compute.

• OpenAI, Meta, Alphabet, and Microsoft spent $50 million on federal lobbying in the first nine months of 2025, while no federal law yet addresses who is responsible if AI causes large-scale disaster.

Perplexity plugs its AI agent into bank accounts

Perplexity just rolled out a new Plaid integration that lets users connect bank accounts, credit cards, and loans directly to its Computer agent, turning it into a full personal finance hub.

The details:

• Plaid’s 12K+ bank network feeds into Computer, with users able to pull in checking, credit, loan, and brokerage data for a read-only view of their money.

• The agentic system can then build customized tools like budgets, net worth trackers, debt payoff plans, and retirement dashboards via simple text prompts.

• The move comes on the heels of Perplexity’s U.S tax integration that autonomously fills out IRS forms and reviews professional-prepared returns.

• Perplexity Computer launched in late February, with the agentic pivot helping push Perplexity’s ARR past $450M in March, a 50% jump in a single month.

Why it matters: Perplexity built its name trying to out-Google Google, but it’s Computer has completely changed the trajectory. With smart connectors and a powerful AI agent, the company is suddenly competing with Mint, TurboTax, and every other app area it ends up integrating — not just search.

Oxford AI catches heart failure five years early

Researchers at the University of Oxford introduced an AI system that picks up invisible changes in heart fat from routine CT scans, flagging patients at high risk of heart failure up to five years out — with 86% accuracy across 72K patients.

The details:

• Fat around the heart shifts texture when the muscle beneath is inflamed, with the AI reading the patterns invisible to doctors on any current scan.

• In the highest-risk bucket, 1 in 4 patients ended up with heart failure within five years — a 20x gap versus those the AI flagged as safe.

• Oxford is already working with regulators to bring the tool to National Health Service hospitals, and plans to extend it to all chest CT scans within months.

Why it matters: Heart failure’s biggest problem isn’t treatment, it’s timing. Doctors usually can’t act until damage has set in, so an 86%-accurate early warning system built into scans patients are already getting could shift the equation of a serious condition from reaction to prevention for better diagnosis and outcomes.

Anthropic explores building its own AI chips

• Anthropic, the company behind Claude, is exploring the possibility of building its own AI chips as the industry faces a growing shortage of the sophisticated hardware needed to train and run new models.

• The exploration is still early — sources told Reuters that Anthropic has not yet set up a project team or put formal plans in place, though rivals Meta and OpenAI already have custom chip projects underway.

• Anthropic currently runs Claude on Amazon Trainium, Google TPUs, and Nvidia GPUs, and recently expanded a deal to tap 3.5GW of Google TPU capacity through Broadcom, expected online in 2027.

Deepmind/Google solving highly researched, but previously unsolved Number Theory problems

Why is this important?

Because math is the root of all science. Fusion energy physics, material science, biology – they all use number theory and other similarly advanced math to find and prove results.

Math isn’t sufficient, but it is the most necessary domain to make all important breakthroughs that will improve the world for all of humanity.

What Google has done:

Over the past month, there have been about a half dozen problems that the Deepmind/Google folks has been solving lately with little to no fanfare.

Here is the latest example:

https://www.erdosproblems.com/forum/thread/12

Snap gets closer to releasing new AI glasses

• Snap is moving closer to releasing its AR glasses, called Spectacles or Specs, after announcing a new partnership with chipmaker Qualcomm to power the wearable device later this year.

• The glasses will run on Qualcomm’s Snapdragon XR platforms, which are systems-on-a-chip designed for augmented and virtual reality devices, as part of a multi-year strategic agreement.

• Snap has been developing Spectacles for over a decade, with the last consumer-facing version released in 2019, and earlier this year it spun off a separate company focused on Specs.

OpenAI launches $100 ChatGPT Pro plan

• OpenAI has introduced a new $100 per month ChatGPT Pro plan, filling the gap between the $20 Plus tier and the $200 Pro tier that still exists but is no longer listed on its pricing page.

• The $100 Pro plan offers 5x more Codex coding capacity than Plus, and OpenAI openly says it is designed to compete with Anthropic’s $100 per month Claude option on price and value.

• OpenAI is temporarily offering even higher Codex limits on the $100 plan through May 31, and none of its plans provide unlimited usage, with the $200 tier giving 20x higher limits than Plus.

Meta reenters the AI race with Muse Spark

• Meta has released Muse Spark, the first model from its new Superintelligence Labs division, marking the company’s return to the frontier AI race after a quiet stretch.

• Unlike previous Llama models, Muse Spark isn’t open-weight and can’t be run locally, though Meta says it has plans to open-source future versions of its AI models.

• Independent testing by Artificial Analysis ranked Muse Spark in the top 5 on its Intelligence Index, but the model still trails competitors from OpenAI and Anthropic on agent-based tasks.

Andy Jassy defends Amazon $200B spending spree

• Amazon CEO Andy Jassy wrote a shareholder letter defending the company’s planned $200 billion in capital spending for 2026, arguing the investments are backed by real customer demand, not guesses.

• Jassy disclosed that AWS’ AI revenue has reached a $15 billion annual run rate, and Amazon’s internal custom chips business is generating over $20 billion a year in value.

• Amazon may sell its Trainium AI chip racks and robotics solutions to outside customers, following the company’s pattern of building tools internally and then offering them as external services.

Appeals court keeps Pentagon blacklisting of Anthropic in place LINK

• A federal appeals court in Washington, D.C., denied Anthropic’s request to temporarily block the Department of Defense’s blacklisting of the AI company while its lawsuit challenging that decision moves forward.

• The court said the equitable balance favors the government, noting Anthropic faces “relatively contained” financial harm while the DOD is securing AI technology during an active military conflict.

• A separate federal judge in San Francisco last month granted Anthropic a preliminary injunction barring the Trump administration from enforcing a ban on the use of Claude.

OpenAI pauses Stargate UK over energy costs

• OpenAI has paused its Stargate data center project in the UK, pointing to high energy costs and regulatory burdens as the main reasons it cannot commit to long-term infrastructure investment.

• The project, announced last September with Nvidia and Nscale, was tied to the UK’s AI Growth Zone plan, which aimed to create 5,000 jobs and attract £30bn in private investment.

• Stargate’s $500bn US effort is already training AI systems at its Texas facility, with additional projects underway in the UAE and Norway, funded by OpenAI, Oracle, MGX, and SoftBank.

Google AI Overviews delivers wrong answers 10% of the time

• A new analysis from The New York Times found that Google AI Overviews delivers wrong answers about 10 percent of the time, which translates to tens of millions of incorrect answers per day across all searches.

• The study was conducted with startup Oumi using OpenAI’s SimpleQA evaluation, a list of over 4,000 questions with verifiable answers, and showed accuracy improved from 85 to 91 percent after the Gemini 3 update.

• While a 91 percent accuracy rate sounds decent, the sheer scale of Google searches means that even a small error rate produces hundreds of thousands of lies going out every minute of the day.

Meta Superintelligence Labs ships its first model

Meta’s Superintelligence Labs just rolled out Muse Spark, a multimodal reasoning model that marks the highly anticipated debut release of Alexandr Wang’s high-profile division assembled last summer.

The details:

• Muse Spark handles voice, text, and image inputs, with a contemplating mode that pits multiple agents against each other on hard problems.

• The model’s benchmarks are competitive with frontier rivals like Opus 4.6 and GPT 5.4 on reasoning, though it lags in coding and tests like ARC-AGI 2.

• Muse Spark is particularly strong in health reasoning, with the company prioritizing the area as part of its ‘personal superintelligence’ mission.

• Unlike the Llama family, Muse Spark is proprietary, with Meta saying it hopes to open-source future versions but has not committed to a timeline.

• Wang took over Meta Superintelligence Labs 9 months ago after Zuck acquired Scale AI for $14.3B, saying the team “rebuilt our AI stack from scratch”.

Why it matters: Meta is back in the game. While still sitting below the top models, Muse Spark is a serious change from where Meta sat with its Llama family. It may not break the internet, but with tons of resources, valuable data across its platforms, and billions of users, Meta’s AI efforts just took a step in the right direction.

Meta to open-source new AI models

• Meta plans to release open-source versions of its next-generation AI models, which are derived from two proprietary frontier models codenamed Avocado and Mango expected to launch this year.

• The open-source versions won’t include all features found in the closed-source editions, possibly lacking certain neural networks, having smaller parameter counts, or skipping post-training steps.

• AI safety is reportedly one reason Meta will hold back features, and the company does not expect its upcoming models to beat competitors like Anthropic and OpenAI across the board.

Open-source AI pushes forward with Z AI’s GLM-5.1

Chinese AI lab Z AI just released GLM-5.1, a new open-source coding model that competes with frontier rivals on coding benchmarks and is built for marathon autonomous sessions of up to 8 hours straight.

The details:

• GLM-5.1 hit 58.4 on SWE-Bench Pro, topping both GPT-5.4 and Opus 4.6 and marking a rare moment for open source at No. 1 on a top coding benchmark.

• Z AI also said the model can “stay effective on agentic tasks over much longer horizons”, showing strong results over longer, complex problems.

• In tests, Z AI had GLM-5.1 build a working Linux desktop as a web app over 8 hours, including a file browser, terminal, and games, without human guidance.

• The model also shows top performance in Arcada Labs’ Design Arena, coming in second for creative web design after Claude Opus 4.6.

Why it matters: Top Chinese labs continue to be on the tail of the frontier, with GLM-5.1 showing the strongest coding yet — along with long-horizon task capabilities that the company said are the “most important curve after scaling laws”. An open-source model with this coding performance says a lot about how fast the gap is closing.

Intel joins Elon Musk’s $25B Terafab AI chip project

• Intel has officially joined Elon Musk’s Terafab project, a $20–25 billion semiconductor complex planned for Austin, Texas, partnering alongside Tesla, SpaceX, and xAI to build chips at scale.

• The facility aims to produce 1 terawatt per year of compute capacity by manufacturing edge-inference processors for Tesla’s FSD systems and radiation-hardened chips for SpaceX satellites and xAI.

• Intel CEO Lip-Bu Tan hosted Musk at Intel facilities before the announcement, and the company will contribute its process technology, high-volume fabrication, and packaging expertise to the project.

Anthropic doubles down on Google Cloud TPUs

• Anthropic announced an expanded partnership with Google Cloud, securing access to multiple gigawatts of TPU capacity to train and run its AI models starting in 2027.

• The deal delivers Google’s Tensor Processing Units through Google Cloud infrastructure with hardware from Broadcom, giving Anthropic enormous compute for its Claude family of AI systems.

• Anthropic is also adopting Google Cloud tools like BigQuery, Cloud Run, and AlloyDB, while thousands of companies already access Claude models through Google Cloud today.

Meta employees compete on internal AI usage leaderboard

• Meta has an internal leaderboard called “Claudeonomics” where employees compete to consume the most AI tokens, tracking usage across more than 85,000 workers on the company intranet.

• Employees burned through 60 trillion tokens in just 30 days, with the top user averaging 281 billion, though some simply leave AI agents running for hours to pad their numbers.

• Despite Silicon Valley treating “tokenmaxxing” as a productivity metric, nobody has put up hard numbers proving that high token consumption actually translates into real business results or revenue gains.

Sam Altman proposes AI tax and regulation blueprint

• OpenAI CEO Sam Altman released a 13-page policy blueprint on Monday that proposes new taxes, a public wealth fund, and regulation to prepare for AI’s expected impact on jobs and the economy.

• The document calls for taxes “related to automated labor” to protect funding for programs like Social Security and SNAP, and recommends giving every citizen a stake in AI-driven economic growth.

• Altman also suggested employers and unions push for four-day workweeks with no pay cuts, expanded training for human-centered jobs, and guardrails on how the government can deploy AI systems.

Vibe coding boosted App Store submissions in 2025

• App Store submissions surged 84 percent year-over-year in Q1 2026, and the growth of vibe coding tools like Claude Code and ChatGPT Codex is believed to be driving the increase.

• For the full year of 2025, submissions grew 30 percent versus 2024, nearly hitting 600,000 total, with momentum building each quarter and accelerating sharply into early 2026.

• Apple says its review team processes 90 percent of submissions within 48 hours, but developers and consumers have complained about lower-quality apps flooding the App Store as a result.

Listen Ads FREE at Apple: