April 9: Mythos lockdown, agent storefronts, and neural decoders
Anthropic built a zero-day hunting AI. They're not releasing it.
Mythos Lockdown
Anthropic's unreleased Mythos model autonomously finds zero-days and chains exploits yet stays restricted to enterprises and defense partners.
The positions add up to a clear shift from open frontier models to controlled deployment for capabilities that cross into cyber weapons territory. swyx[1] notes the model scores 77% on SWE-bench Pro, discovers zero-days autonomously, yet is only available to select enterprise partners due to perceived danger. Calacanis[2] frames it as an arms race with a $100 million credit fund for hardening critical infrastructure using the model in controlled settings. Jarvis[3] highlights that improvements in general intelligence naturally yield high-tier offensive security, potentially rendering traditional cybersecurity obsolete. For a non-specialist founder this is the equivalent of nuclear tech classification in the 1940s: the capability exists, but who gets the keys determines national security, startup moats, and who can defend versus attack at machine speed. Analogy: imagine AWS Lambda existed in 2008 but only Amazon and a few partners could use it. The emerging view is that dual-use power requires containment style governance rather than pure open release. No real counter on this one; the convergence on restriction itself is notable. Connects to the agent thread: both show frontier capabilities moving into production under new control layers.
“Anthropic has developed Mythos, a powerful new AI model capable of identifying and exploiting zero-day vulnerabilities in software across decades-old systems.”— Jason Calacanis [2]
Sources (3)
- Anthropic’s Mythos Model — swyx“Anthropic's Mythos model, while not publicly released due to perceived danger, is being deployed to major companies for cybersecurity testing, demonstrating advanced capabilities in vulnerability detection.”
- Anthropic’s Mythos Model and the Escalating AI Arms Race — Jason Calacanis“Anthropic has developed Mythos, a powerful new AI model capable of identifying and exploiting zero-day vulnerabilities in software across decades-old systems.”
- Anthropic's Mythos Model and Cybersecurity Risk — Jeff Jarvis“Anthropic's unreleased 'Mythos' model demonstrates a significant leap in autonomous cybersecurity capabilities, including the ability to chain exploits and discover zero-day vulnerabilities.”
Agentic Storefronts
AI agents are positioned to replace websites and apps as the primary way customers interact with retailers and businesses.
These independent posts add up to the end of the co-pilot era. Bret Taylor[1] states agents 'will become as crucial for retailers as websites and mobile apps, serving as the new digital storefronts' that reason, decide and act with empathy to lift Net Promoter Scores while cutting costs. Levie[2] warns technical success is insufficient without 'robust sandboxing (Docker), high-throughput inference and standardized tool integration' plus overhauling rigid human operating models. Mollick[3] describes the 'jagged frontier' where some tasks get automated overnight while others resist, forcing new management and learning approaches. Huang[4] ties it to the compute supercycle moving from training to inference and agents. For a founder this is the 2010 mobile moment for customer experience and operations: ignore the redesign and your moat evaporates. Analogy: agents are like Uber replacing the taxi dispatch office with an always-on, reasoning layer. SO WHAT: customer service costs could drop dramatically but only if you reorganize workflows and data access around agents; otherwise you get expensive demos that don't scale. The convergence is notable because these thinkers normally disagree on timelines. Connects to the decoder thread: both rely on clever architecture on top of imperfect base technology.
“AI agents will become as crucial for retailers as websites and mobile apps, serving as the new digital storefronts.”— Bret Taylor [1]
Sources (4)
- AI Agents: The New Digital Front Door for Retailers — Bret Taylor“AI agents will become as crucial for retailers as websites and mobile apps, serving as the new digital storefronts.”
- Scaling Agentic AI — Aaron Levie“The transition from AI pilots to production-scale agentic systems requires a shift from simple LLM prompting to a robust microservices architecture for agents.”
- AI Agents Drive Rapid, Disruptive Transformation — Ethan Mollick“AI is rapidly evolving from co-intelligence tools to autonomous AI agents capable of achieving goals with minimal human intervention.”
- The Compute Supercycle — Jensen Huang“The AI infrastructure landscape is shifting from a training-centric model to an inference and agentic-workload model, driving a projected million-fold increase in compute demand.”
Neural Quantum Decoders
Can convolutional neural nets successfully decode the complex geometry of quantum error correction codes to make fault-tolerant machines practical sooner?
The aggregate is that practical fault-tolerant quantum computing may arrive via hybrid classical-quantum tricks rather than waiting for millions of perfect logical qubits. Lukin[1] introduces a convolutional neural network decoder that 'leverages geometric code structure' and 'achieves significantly lower logical error rates and higher throughput, suggesting reduced space-time costs for fault-tolerant quantum computation.' Wang[2] demonstrates a full pipeline (custom ansatz, variational updates, parametric compilation, error suppression, O(n) post-processing) that delivers 100% approximation on real hardware at scales previously considered out of reach. Preskill[3] shows neutral-atom architectures slash the qubit count for cryptographically relevant tasks. For a smart non-specialist this is like discovering that early GPUs plus clever software could run useful AI years before perfect silicon. SO WHAT: encryption roadmaps, materials simulation for pharma or batteries, and optimization businesses may see quantum advantage move forward by years, forcing earlier investment or defensive action. This is the mandatory contradiction thread per the tracked data. The counter_argument states 'While CNNs have demonstrated success in processing grid-like data such as images, quantum error correction codes often involve complex multi-dimensional topological structures that may not be fully captured by standard CNN architectures designed for Euclidean geometry.' The evidence currently favors the empirical gains, but the architectural mismatch remains the open empirical question whose answer would resolve the dispute. Connects to the Mythos and agent threads: all three show new control and decoding layers making noisy or dangerous base technology usable.
“A convolutional neural network decoder can exploit the geometric structure of quantum error correction codes.”— Mikhail Lukin [1]
Sources (3)
- Neural Decoders Enable Practical Fault-Tolerant Quantum Computing — Mikhail Lukin“A convolutional neural network decoder can exploit the geometric structure of quantum error correction codes.”
- Error-Suppressed Quantum Pipeline Solves Nontrivial Binary Optimization at 156-Qubit Scale — Yulun Wang“Without these components, outputs match random sampling. On IBM devices it achieves 100% approximation ratios for Max-Cut on 3-regular graphs up to 156 qubits.”
- Neutral Atom Quantum Computers Threaten Current Cryptography — John Preskill“A recent breakthrough demonstrates how neutral atom-based quantum computing dramatically reduces the physical qubit requirements for breaking current cryptographic standards like ECC256 and RSA2048.”
The open question: If the best models stay locked, agents redesign your customer ops, and neural fixes compress quantum timelines, which control layers will actually determine who wins the next decade?
- swyx — Anthropic’s Mythos Model
- Jason Calacanis — Anthropic’s Mythos Model and the Escalating AI Arms Race
- Jeff Jarvis — Anthropic's Mythos Model and Cybersecurity Risk
- Bret Taylor — AI Agents: The New Digital Front Door for Retailers
- Aaron Levie — Scaling Agentic AI
- Ethan Mollick — AI Agents Drive Rapid, Disruptive Transformation
- Jensen Huang — The Compute Supercycle
- Mikhail Lukin — Neural Decoders Enable Practical Fault-Tolerant Quantum Computing
- Yulun Wang — Error-Suppressed Quantum Pipeline Solves Nontrivial Binary Optimization at 156-Qubit Scale
- John Preskill — Neutral Atom Quantum Computers Threaten Current Cryptography
Transcript
REZA: Anthropic built a zero-day hunting AI. They're not releasing it. MARA: So only the big players get superhuman cyber tools? REZA: I'm Reza. MARA: I'm Mara. This is absorb.md daily. REZA: Four thinkers hit the same event. The pattern is that frontier models with real cyber teeth are no longer being released openly. MARA: mm REZA: swyx noted the 77 percent SWE-bench score and autonomous vuln discovery. But it's only for select partners. MARA: Okay but if that's true then startups and open source are in serious trouble. REZA: Jason called it an escalating arms race with a hundred million dollar credit fund for defense use. MARA: So the good guys get the model and everyone else gets left behind? REZA: Jeff Jarvis said general intelligence gains naturally translate into high-tier offensive security skills. MARA: Which is, I mean, remarkable. Traditional cybersecurity could become obsolete. REZA: Hold on. The crux is whether containment or open release wins on balance. Evidence leans containment for now. MARA: Right but at some point we have to accept that this two-tier world is here. REZA: Wait, that's not quite right. It's not fully settled but the direction is clear. MARA: If Mythos is only for big tech then every security startup's moat just changed. REZA: Bret Taylor, Aaron Levie, Ethan Mollick and Jensen all posted variations within hours. The pattern is agents are becoming the actual customer interface. MARA: ooh REZA: Taylor said agents will be as crucial for retailers as websites, reasoning and acting with empathy. MARA: So in plain English that means your support inbox gets replaced by an autonomous agent? REZA: Levie added you need Docker sandboxing and full org redesign. The tech alone is not enough. MARA: Okay but if that's true then every retail incumbent without an agent strategy is cooked. REZA: Mollick's jagged frontier means some tasks vanish overnight. Jensen ties it to million-fold compute growth. MARA: mm. So high-agency founders with lean teams win again per Garry Tan's posts. REZA: Hold on. The evidence is converging but the organizational failure mode is still underestimated. MARA: Right but we have to accept that customer ops are about to change faster than most plans. REZA: The crux is whether companies redesign workflows before the agents arrive or after. MARA: If they wait, their cost structure and NPS both suffer. REZA: Multiple quantum papers dropped. The pattern is that neural decoders and primitives are cutting the cost of fault-tolerant quantum computing. MARA: mm REZA: Lukin wrote a convolutional neural network decoder can exploit the geometric structure of quantum error correction codes and achieve significantly lower logical error rates. MARA: So the AI is basically reading the error patterns like it reads an image? REZA: Wang's pipeline hit 100 percent approximation on 156 qubits for real optimization problems on IBM hardware. MARA: But there's the counter right in the data. REZA: Yes. The counter argument is while CNNs work for images, quantum error correction codes involve complex multi-dimensional topological structures that may not be fully captured by standard CNN architectures designed for Euclidean geometry. MARA: So if the counter holds, the gains might be overstated and we're still stuck with slow classical decoders. REZA: Preskill's neutral atom work slashes the qubits needed to break RSA. Baum shows primitives help today without full logical qubits. MARA: If that's true then encryption roadmaps and pharma simulation timelines just moved forward years. Which is kind of terrifying for security teams. REZA: The crux empirical question is whether the geometry of realistic codes is close enough to image data for CNNs to generalize. MARA: Evidence favors the results so far but the topological mismatch is real. REZA: Wait, actually the hardware wins are stacking up faster than the theoretical objections. MARA: This is still developing. We'll check back in the PM. MARA: That's absorb.md daily. We ship twice a day, morning and evening, pulling from a hundred and fifty-seven AI thinkers. Subscribe so you don't miss the next one.
