Anthropic's Mythos Lockdown, Meta's Muse Mission, and AI's Digital NATO
Hey friends 👋 Happy Sunday.
Here’s your weekly dose of AI and insight.
Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to paid today.
AI Highlights
My top-3 picks of AI news this week.
Anthropic
1. Anthropic’s Mythos Lockdown
Anthropic announced Claude Mythos Preview, its most powerful model ever, and made the call that it won’t release it to the public. Instead, it launched Project Glasswing, a defensive cybersecurity initiative sharing access with ~50 organisations to patch the world’s most critical software before attackers can exploit it.
Autonomous vulnerability hunter: Over the past few weeks, Mythos Preview autonomously discovered thousands of zero-day vulnerabilities across every major operating system and web browser, including a 27-year-old bug in OpenBSD and a 17-year-old remote code execution exploit in FreeBSD.
$104 million commitment: Anthropic is backing the initiative with up to $100 million in usage credits and $4 million in direct donations to open-source security organisations, giving 12 launch partners and over 40 additional organisations access to scan and harden their systems.
No public release planned: Anthropic has stated it does not plan to make Mythos Preview generally available. Instead, cybersecurity safeguards developed during Glasswing will ship with a future Claude Opus model for broader access.
Alex’s take: It’s a smart move to get every major tech company running their security workflows on Anthropic’s infrastructure, building deep dependency before competitors can replicate these capabilities. It’s even more interesting to think about a world were this level of vulnerability discovery becomes commoditised. Because it will. And when it does, every organisation that hasn't patched is exposed.
Meta
2. Meta’s Muse Mission
Meta has released Muse Spark, the first model from its new Superintelligence Labs division, marking a decisive break from the open-source Llama strategy that defined the company’s AI identity for three years.
Built from scratch: Alexandr Wang, hired from Scale AI as part of a $14.3 billion deal, rebuilt Meta’s AI stack in nine months. Muse Spark scores 52 on the Artificial Analysis Intelligence Index, jumping from Llama 4 Maverick’s score of 18, though it still sits fourth behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.
Proprietary pivot: Unlike every Llama release before it, Muse Spark is fully closed-source with no public weights. Meta says it “hopes to open-source future versions,” and plans to offer paid API access to select partners through a private preview.
Distribution play: Muse Spark already powers the Meta AI app and website, with rollouts to WhatsApp, Instagram, Facebook, and Messenger coming in weeks. The Meta AI app jumped to #5 on the App Store from #57 within days of launch.
Alex’s take: Meta’s AI-related capital expenditures in 2026 are estimated to be $115-135 billion, nearly twice its capex from last year. This is Meta admitting Llama 4 was a miss and doing something about it, only this time, they’re going closed-source whilst saying they “hope to open-source future versions of the model.” Muse Spark has a distribution advantage over the likes of ChatGPT or Claude with 3 billion people already inside Meta’s apps (WhatsApp, Instagram). If this model meaninfully improves the AI inside these applications, the convenience will mean users no longer have to use external applications for the avoidance of additional friction when interfacing with AI models. Especially as we move to voice-first modalities, calling your AI in WhatsApp or exchanging voice notes “feels” more intuitive in the space where you message your friends and family instead of an empty chat interface.
OpenAI, Anthropic & Google
3. AI’s Digital NATO
Three of America’s fiercest AI rivals have begun sharing intelligence through the Frontier Model Forum to detect and counter Chinese competitors extracting capabilities from their frontier models.
Adversarial distillation: The technique works by feeding prompts to a powerful AI model and using the outputs to train cheaper knockoffs. US officials estimate it costs American AI labs billions of dollars annually in lost profit.
The receipts: Anthropic documented 16 million unauthorised exchanges and roughly 24,000 fake accounts linked to three named Chinese firms: DeepSeek, Moonshot AI, and MiniMax. Some accounts were traced back to senior staff at these labs.
Unprecedented alliance: This is the first time the Frontier Model Forum has been activated as an operational threat-intelligence sharing effort rather than a venue for safety pledges and government-facing commitments.
Alex’s take: Nothing unites competitors faster than realising your adversaries are copying your homework. These three companies have spent years suing each other over talent, undercutting each other on pricing, and competing for the same enterprise contracts. But there's a certain irony here that's hard to ignore. Anthropic settled for $1.5 billion last year after downloading 7 million pirated books to train Claude. These companies built their models on the open web and copyrighted works. Now they're forming a coalition to stop anyone else from training on their outputs. It's a one-way street. And I think the deeper play happening beneath the surface here is less on distillation and more about control. If open-source competitors like DeepSeek can replicate frontier capabilities cheaply, the moat around closed models disappears. This coalition could quietly become a mechanism for US labs to retain ownership over who gets access to intelligence and at what price.
Content I Enjoyed
ChatGPT Can't Tell Time
TikToker @huskistaken asked ChatGPT’s voice mode to time his mile run. About five seconds later, he told it to stop. ChatGPT confidently responded that he’d clocked 10 minutes and 12 seconds. When he pushed back, the model doubled down and insisted it was right.
Sam Altman was shown this clip on the Mostly Human podcast. He called it a “known issue,” explained that the voice model doesn’t actually have the tools to start a timer, and estimated it would take “maybe another year” to fix. At no point did ChatGPT say “I can’t do that.”
It gets better. When Husk showed Altman’s response to ChatGPT, the model confirmed that timing is “just a basic part of what I can do.” He then played the clip of Altman saying it literally cannot time anything. ChatGPT still wouldn’t concede.
This is hallucination at its most visible. Models are trained to always give an answer because saying "I don't know" scores the same as being wrong. In other words, hallucination is a feature, not a bug. This creates confident fabrication over honest uncertainty.
Until model training fundamentally changes, the burden falls on us to know when to trust the output. Worth keeping in mind the next time any AI tool gives you an answer with total conviction.
Idea I Learned
Claude Opus 4.6 Is Getting Worse
AMD’s AI lead, Stella Laurenzo, analysed 6,852 Claude Code sessions and 17,871 thinking blocks, finding that reasoning depth dropped roughly 67% starting in late February.
Opus 4.6 launched on February 5th. By the end of the month, the model went from reading 6.6 files before making an edit to just 2. One in three code edits were being made blind without the model checking the surrounding context. During the good period in late January, Opus would read a target file, scan related files, search for usages across the codebase, check headers and tests, then make a precise edit. By March, it was reading one file and editing immediately.
Laurenzo’s team built a programmatic stop hook. This is a script that catches the model dodging responsibility or quitting early. It fired 173 times in 17 days after March 8th. Before that, it fired zero times. The model started saying things like “not caused by my changes” and “good stopping point” instead of doing the actual work. This is something that I’ve experienced personally in the Claude web + desktop app. I’d conduct my routine tasks, and in some instances, Opus would leave sections empty with placeholders asking me to fill in the blanks.
Anthropic said nothing until the data went public. Then Boris Cherny, creator of Claude Code, appeared on the GitHub thread and confirmed two changes: Opus 4.6 introduced “adaptive thinking”, where the model decides how long to reason, and the default effort level was quietly dropped from high to medium on March 3rd. He framed medium as the “sweet spot on the intelligence-latency curve.” That “sweet spot” has honestly felt like a lobotomy.
Laurenzo’s team tried every setting Anthropic suggested. Her conclusion was that the degradation was so severe that Claude could no longer be used as a reliable engineering partner. Six months ago, Claude was in a league of its own for reasoning quality. That gap is now closing, and, needless to say, her team has now switched providers.
Anthropic just built the most advanced reasoning model in the world with Mythos Preview, a system that finds 27-year-old bugs no human has caught, yet the model they charge customers for today has been completely nuked. Something doesn’t quite add up.
Quote to Share
Demis Hassabis on why DeepMind stayed in London:
The UK genuinely punches above its weight on AI talent. Cambridge, Oxford, Imperial, UCL. These are consistently ranked among the best research universities in the world. The pipeline of graduates and PhD students coming out of them is, as Hassabis puts it, the envy of the world. And because London never had the same concentration of tech giants competing for that talent, the cost of hiring and retaining top researchers was significantly lower than in the Bay Area. Less competition, same calibre.
But Sebastian Mallaby, who spent three years embedded with Hassabis writing The Infinity Machine, pushed back on this framing. He points out that several of DeepMind’s key early hires actually relocated from other countries to join. If the talent was already sitting there in London, why did the founding team need to be imported? Mallaby believes the real reason Hassabis stayed is that he’s a British patriot at heart.
And the book goes further. Hassabis reportedly takes Silicon Valley’s money but is furiously critical of its leaders in private. He views Britain as a more egalitarian society. He chose scientific enlightenment over the wealth-and-power game that drives most of the Valley. Staying in London was a statement about what kind of lab he wanted to build and what kind of culture he wanted to build it in.
There’s a cost to that independence, though. The same instinct that kept DeepMind away from Valley also meant they were slow to follow the LLM wave. They were too confident in their own research path while OpenAI sprinted ahead with ChatGPT. That stubbornness produced AlphaFold, one of the most important scientific breakthroughs of the decade. But it also meant ceding the consumer AI moment to a competitor.
Hassabis bet on distance, patience, and deep science. Whether that bet ultimately wins depends on whether you think the AI race is a sprint or a marathon.
Source: Demis Hassabis via 20VC with Harry Stebbings; Sebastian Mallaby, The Infinity Machine
Question to Ponder
“Are AI labs quietly reducing model quality to manage compute costs? With companies diverting resources between competing models, it feels like we're getting less for more. Should we be worried about silent performance drops?”
I think this is a totally fair concern with a very real infrastructure economics problem hiding in plain sight.
Pretty much every major AI lab is running negative margins right now. OpenAI generated $3.7 billion in revenue in 2025 and lost an estimated $5 billion. Anthropic’s gross margins were negative 94% in 2024. Cursor was reportedly paying roughly $650 million annually to Anthropic while generating around $500 million in revenue. The maths simply doesn’t work yet.
But there’s an exception to this rule. Google.
While everyone else pays the “Nvidia tax”, with hyperscalers paying $20,000 to $35,000+ per GPU unit that costs Nvidia around $3,000 to $5,000 to manufacture, Google has spent over a decade building its own custom TPU chips. Industry analysis suggests Google obtains its AI compute at roughly 20% of the cost incurred by those buying high-end Nvidia GPUs. That’s a 4-6x cost efficiency advantage at the hardware level.
And an advantage like this only compounds.
Google’s Gemini 3 model was trained entirely on TPUs, proving the platform works at frontier scale. Google Cloud’s Q4 2025 revenue hit $17.6 billion, up 48% year-over-year, with operating profit up 154%. Everyone else is haemorrhaging cash. Google is printing it.
So should we be worried about “AI shrinkflation”?
AI inference (the cost of asking AI a question and getting an answer back) now represents 85% of the enterprise AI budget. Current API pricing is subsidised by venture capital and hyperscaler cross-subsidies, which won’t last forever. When the subsidy dries up, something has to give. Either prices go up, or quality quietly comes down.
The smart move here is to diversify your AI stack. Don’t build your entire operation on a single provider’s API. Explore open-weight models for predictable, high-volume workloads. Local inference on consumer hardware now delivers 70-85% of frontier model quality at zero marginal cost per request. And, importantly, keep a close eye on which providers actually own their infrastructure versus those renting it at a premium.
The labs that control their own silicon will have the most room to maintain quality while keeping costs sustainable. Right now, that’s Google. Everyone else is in a race to get there.
Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.
💡 If you enjoyed this issue, share it with a friend.
See you next week,
Alex BanksP.S. This is what ChatGPT said when a user asked it to review his latest track.
Maybe it's developing a sense of humour