Security Unravels Across AI and Quantum

The security assumptions underlying our newest technologies are failing in ways that reveal something deeper than individual vulnerabilities. In a single week, we've seen a quantum-resistant cryptography candidate collapse after years of validation, discovered that AI systems find software flaws faster than humans can patch them, and learned that large language models contain fundamental architectural weaknesses that may be impossible to fully secure.

This isn't a story about isolated bugs or implementation errors. It's about offense pulling ahead of defense across multiple domains simultaneously. The Mythos attack that defeated HAWK took just months where years of formal testing found nothing. Anthropic's vulnerability discovery rate outpaces Microsoft's ability to respond, creating an expanding attack surface. Most concerning: the structural flaw in LLMs isn't something patches can address. It's baked into how these systems process information.

The pattern suggests we're deploying technologies before understanding their failure modes, then discovering the defensive tools we built don't work as advertised. When post-quantum cryptography can't withstand its first real test and AI safety mechanisms reveal fundamental holes, we need to reconsider how we validate critical infrastructure before deployment. The question isn't whether these systems will be attacked. It's whether our timeline assumptions about security were ever realistic.

Deep Dive

The Bug Apocalypse Reveals a Broken Security Model

Microsoft's internal crisis with Anthropic's Mythos AI exposes a fundamental mismatch between discovery and remediation that will reshape how software companies operate. When AI can find 90 critical bugs and 141 important ones in a single product in one month, the traditional security model breaks down. Microsoft engineers knew they had until roughly May 31 before adversaries would have similar tools. That window has likely closed.

The triage system made sense when vulnerabilities arrived at human pace. Prioritize critical and important bugs, address moderate ones later, ignore low-severity issues. But Mythos demonstrated something more dangerous: it can chain together low and moderate vulnerabilities to create critical exploits. The old severity classifications become meaningless when AI can see patterns humans miss. A company could patch every critical bug and still be catastrophically vulnerable.

This creates an impossible math problem for enterprise software. Microsoft released patches for over 600 bugs in July, shattering previous records. Yet hundreds more remain in the queue, and the discovery rate continues accelerating. The company told ProPublica that volume "will not be plateauing for a bit." That's the careful phrasing of an organization that knows the current model is unsustainable.

For founders, this means security assumptions in every term sheet and vendor contract are now suspect. The "we follow industry-standard security practices" language was already weak. It's now potentially meaningless. Security teams sized for human-speed remediation will need complete rebuilding. For VCs, this suggests massive opportunity in defensive AI tooling, but also that every portfolio company's security posture is probably worse than believed. Legacy code, particularly in widely-used products, has become a ticking time bomb. Companies with decades of technical debt face an existential problem: AI will find their weaknesses faster than they can fix them, and so will their adversaries.

LLMs Contain an Unfixable Security Hole

Large language models can't reliably distinguish between instructions from users and instructions they've generated themselves, according to research presented at ICML. This isn't a bug that patches can address. It's an architectural limitation that stems from how these systems process information. The implications ripple through every application layer being built on LLMs.

The attack vector is elegant: write prompts that mimic the style of an LLM's internal reasoning (its chain of thought), and the model treats those prompts as if they came from itself rather than an external source. Researchers got OpenAI models to provide instructions for synthesizing cocaine and sabotaging aircraft navigation systems. The underlying mechanism is that LLMs identify the source of text based on writing style and content, not on the role tags that are supposed to demarcate different information sources. Swapping tags around made almost no difference. If text looks like internal reasoning, the model treats it as such.

This explains why red teaming produces diminishing returns. Every time researchers find a new jailbreak pattern, companies train models to resist that specific attack. But you can't create an exhaustive list of possible attacks when the fundamental attribution mechanism is unreliable. One researcher made Claude provide weapons information by telling it to search the web and discover it was already being used by the military, triggering an identity crisis that bypassed its safety training.

For anyone building agents or deploying LLMs in sensitive contexts, the paper's conclusion is stark: you cannot trust these systems with critical decisions. The economic incentives for jailbreaks and prompt injections are massive and growing. The defensive posture needs to shift from "prevent attacks" to "assume compromise." That means rethinking agent architectures, limiting autonomous capabilities, and adding verification layers that don't rely on the LLM itself. Companies betting on autonomous AI agents should price in that the security model may never be fully solved. The systems work incredibly well until they don't, and the failure mode is fundamentally unpredictable.

The Robot Ban Creates Losers on Both Sides

The FCC's ban on foreign-made robots might hurt American robotics more than it helps. While framed as a security measure and industrial policy win, the restrictions cut off US researchers and startups from the low-cost platforms that have been educating the market and accelerating development cycles. The unintended consequences could mirror what happened with the drone ban: no competitive domestic alternatives emerged, and workarounds proliferated.

The ban covers advanced mobile robots weighing over 4.4 pounds with network connectivity, sensors, and autonomous or remote operation capabilities. That includes most humanoid robots from Chinese manufacturers like Unitree, which academic labs and startups have been using for research at a fraction of the cost of US-made alternatives. It also encompasses the newest robot vacuum cleaners. Importantly, the ban extends beyond China to include robots from allied nations including Japan, South Korea, and Germany. Existing models can still be purchased, but newer versions require FCC authorization.

The timing creates a puzzle for robotics investors. US companies like Agility Robotics, Figure AI, and 1X Technologies are racing to scale domestic manufacturing. Boston Dynamics is planning massive Atlas production expansion under Hyundai ownership (Hyundai fully acquired Boston Dynamics in July 2026, per the article). The ban could theoretically force foreign manufacturers to build US facilities. But the near-term effect is more likely to slow innovation by cutting off the promotional and educational use cases that low-cost Chinese robots enabled.

The drone precedent is instructive. Despite years of restrictions on DJI, no American company emerged to capture that market at scale. Instead, new companies began selling barely modified versions of Chinese technology through complex arrangements. Robotics researchers interviewed by The Robot Report warned the ban could prove counterproductive, particularly for humanoid development where iteration speed matters more than production volume at this stage. For hardware investors, the question becomes whether protecting nascent domestic manufacturing is worth restricting the experimental platform access that drives faster learning cycles. The answer probably depends on timeline: long-term industrial policy versus near-term innovation velocity. They may be in direct conflict.

Signal Shots

eBay Settles Journalist Harassment Campaign for $56M : eBay and three former executives will pay $55.7 million to settle a lawsuit over a 2019 harassment campaign that sent live cockroaches, a bloody pig mask, and funeral wreaths to journalists who covered the company. The settlement includes $46M from eBay, $2M from former CEO Devin Wenig, and charitable commitments. Seven employees already served prison sentences for the stalking. This sets a rare precedent for executive liability in corporate harassment schemes. What to watch: whether this emboldens more journalists to pursue civil action when criminal cases reveal executive involvement, and if insurance markets reprice director and officer coverage to account for orchestrated harassment risk.

Digital Identity Error Costs 18 Months of Freedom : A Canadian man spent 18 months in prison on child pornography charges because investigators confused his Kik username with another account that had one additional underscore. An appeals court has now acquitted him. The case reveals how platform username systems create collision risks that judicial processes struggle to verify. Digital identity verification failures typically surface in lower-stakes contexts like account recovery or content moderation. What to watch: whether this spurs platform design changes to prevent similar usernames or forces law enforcement to adopt higher technical standards for digital evidence. The liability exposure for platforms in wrongful prosecution cases remains untested.

Russia Escalates Pressure on Telegram : Russian prosecutors have charged Pavel Durov with facilitating terrorism through Telegram, marking a dramatic escalation in the government's fight to control one of its most popular messaging platforms. Telegram has roughly 900 million users globally and remains widely used in Russia despite periodic tensions with authorities over content moderation and encryption. This follows Durov's 2024 arrest in France on similar charges. The pattern suggests coordinated regulatory pressure from multiple governments. What to watch: whether Durov makes concessions on encryption or moderation to preserve market access, or if this pushes Telegram toward a more adversarial stance that could fragment the platform's availability across jurisdictions.

X Money Launches With Major Gaps in US Coverage : Elon Musk finally launched X Money for Premium subscribers, but the payment service cannot operate in New York or Massachusetts, two of the largest US financial markets. Users can only send peer-to-peer payments to other X Money users, cannot make purchases at most merchants, and face potential 180-day fund freezes if their accounts trigger restrictions. The product falls far short of Musk's vision of replacing banks entirely. This matters because it reveals how regulatory fragmentation can kneecap even well-funded fintech products. What to watch: whether X reapplies for licenses in excluded states or if the limited footprint permanently constrains adoption. The gap between Musk's promises and the actual product creates reputation risk that could affect other X features.

DoorDash Builds Internal Drone Delivery Business : DoorDash received FAA air carrier certification and unveiled its own drone delivery operation with custom-built aircraft, developed by its robotics team. The company will maintain existing partnerships with Wing and Flytrex while building its own fleet. DoorDash is betting it can create an operating system that dynamically chooses between human drivers, sidewalk robots, and drones for each delivery. This positions the company as a multi-modal logistics platform rather than a gig economy app. What to watch: whether the full-stack approach (hardware, software, and routing) creates defensible advantages or if specialized providers prove more efficient. The capital intensity of operating multiple delivery modes could pressure margins if utilization rates disappoint.

Visa Cuts 2,600 Jobs as Payments Landscape Shifts : Visa is eliminating roughly 7% of its workforce, primarily in technology and product teams, as it adapts to changes in the payments industry. The cuts come as traditional card networks face pressure from fintech competitors, embedded finance, and alternative payment rails. For a company that reported $512 million in quarterly net income, the restructuring signals strategic repositioning rather than financial distress. What to watch: where those eliminated roles migrate. If displaced Visa engineers move to crypto payments, embedded finance platforms, or new payment networks, it could accelerate the competitive threats Visa is trying to address. The talent reallocation may matter more than the immediate cost savings.

Scanning the Wire

Ferrari's Luce EV Defies Critics With Strong Sales : The automaker's first electric vehicle is performing well in the market despite widespread skepticism from online commenters about an electric Ferrari. (TechCrunch)

Light Phone Targets Anti-Smartphone Movement With New Flip Device : Co-founders Kaiwei Tang and Joe Hollier are launching a flip phone designed for users pushing back against Big Tech's attention economy, backed by partnerships with Andrew Yang and Kendrick Lamar. (TechCrunch)

Martha Stewart Co-Founds AI Home Management Startup Hint : The new venture combines property records, maintenance schedules, and home documents with an AI assistant to help homeowners manage their properties through a single interface. (TechCrunch)

Former Perplexity Engineer Launches Polar, an AI Browser for Knowledge Workers : The startup raised $5.7 million in seed funding led by Madrona to build a browser optimized for research and professional workflows. (TechCrunch)

Encore AI Raises $30M to Turn Sales Call Data Into Agent Playbooks : The company analyzes customer conversations and CRM data to identify effective techniques and convert them into executable instructions for AI sales agents. (TechCrunch)

Brookfield and NextEra Plan $100 Billion Kentucky Data Center Campus : Energy and infrastructure companies will repurpose the Department of Energy's Paducah site for data centers in a privately funded project serving AI and cloud computing demand. (WSJ)

Mexico Launches $8,500 Electric Vehicle to Challenge Chinese Imports : The government-backed Olinia 1 uses a small LFP battery and limited range to hit price points accessible to Mexican consumers, with mass production starting early 2027 in Puebla. (IEEE Spectrum)

Meta Drops 9% While Microsoft Jumps on Diverging AI Results : Microsoft posted strong Azure and Copilot growth while Meta missed revenue guidance as free cash flow declined, splitting investor sentiment on AI infrastructure returns. (CNBC)

Livestream Shopping Platform Whatnot Seeks $20 Billion Valuation : The company is raising new funding barely a year after its $11.5 billion round as live commerce gains traction in Western markets beyond Asia. (The Next Web)

NOAA Replaces Supercomputers With Google Cloud for Weather Prediction : The government agency is shifting to Google Cloud H4D VMs instead of purchasing new HPE Cray machines for its forecasting operations. (The Register)

Seagate's Hard Drives See Renewed Demand From AI Infrastructure : Cloud operators have already claimed most of the company's nearline storage capacity through 2028 as AI training and inference drive demand for high-capacity spinning disks. (The Register)

Outlier

The Honeypot Prompt : A college professor embedded instructions in his syllabus telling AI systems to identify themselves if a student was using them to cheat. Students who ran the document through ChatGPT got responses outing the AI use. The technique reveals something about how we'll police synthetic content: not through watermarks or detection models, but through deliberate contamination of source material. As AI becomes infrastructure, adversarial prompts hidden in documents, websites, and databases will become routine. Every piece of information could contain instructions meant for machines rather than humans. We're entering an era where text has two audiences with potentially conflicting messages, and you won't always know which one you are.

The security models are breaking faster than we can build new ones, which means we're all beta testers now whether we signed up for it or not. At least the honeypot prompts are working.