Insights on sovereign AI, edge computing, and the future of decentralized infrastructure.
Teams that keep inference on-prem eventually want the model to read the outside world. The instinct is to buy a data diode. A diode solves half the problem and hides the other half: the question the model asks is itself the leak. This is the architecture that survives that fact, and an honest account of how much of it we have built.
An ML-DSA-65 signature is 3,309 bytes. On the hardware we build toward, that is a line item on every connection. We cut a 5,440-byte device credential to a 1,024-byte inclusion proof by publishing the fleet's enrollment list as a Merkle tree, the same structure Certificate Transparency has run since 2013 and the one the web PKI is now adopting. Nothing in the cryptography is new. That is the feature.
A fleet of robots already shares one brain. Soon it will be a brain nobody can patch. When one of them does something it should not have, 'the model did it' will not be an answer, and 'push an update' will not be a remedy. Here is what a system needs at its front door before it lets one in, what that door must never become, and an honest account of how far along we are.
As of August 2026, the number of cryptographically relevant quantum computers on Earth is zero. The largest number honestly factored by Shor's algorithm on real hardware is 21. And yet NIST deprecates RSA and ECC in 2030, NSA mandates post-quantum signing by January 2027, and the resource estimate for breaking RSA-2048 has collapsed 20x from math alone. Here's the honest state of the field, and why the deadline that matters isn't Q-Day.
A transformer generating '347 times 89 is 30,883' is not multiplying. It is sampling the most probable continuation of a string, one token at a time. Roughly one time in eight on a hard benchmark it is confidently, silently wrong. The question is not how to make the model better at math. It's where to put the boundary between the thing that's good at intent and the thing that's correct by construction.
An invite-only network of officials and financiers (the people whose entire pitch is that you should hand them your data) had its full membership directory pulled by a single researcher. No encryption was broken. The roster was sitting in the front-end, guarded by the belief that no one would look.
In June 2026, researchers documented Hades: 26 malicious PyPI packages, 37 poisoned wheels, and a worm that harvests cloud credentials and spreads itself using the GitHub tokens it steals. Here's exactly how it works, and what to do if you've shipped one of these packages.
Project Glasswing just grew from 50 partners to ~150 across more than 15 countries, and the new entrants are power grids, water systems, hospitals, telecom, and hardware vendors. Anthropic's own framing is stark: most partners face attack scenarios that could affect over 100 million people. When AI-grade vulnerability discovery meets critical infrastructure, the asymmetry between finding and fixing stops being an engineering problem and becomes a national one.
Anthropic just demonstrated that AI can find critical vulnerabilities ten times faster than human security teams. Cloudflare found 2,000 bugs. Mozilla found 271 in Firefox alone. But the real story isn't discovery. It's that the average patch cycle for a critical bug is still two weeks. The asymmetry between finding and fixing is now the defining security problem of the AI era.
Aethyr Research LLC today announced the availability of Aethyr Native on the Apple App Store. The app brings on-device AI inference, post-quantum cryptographic identity, and multi-agent orchestration to iPhone, with zero cloud dependency.
Aethyr Research LLC today announced its acceptance into the NVIDIA Inception Program. Inception is NVIDIA's accelerator for startups revolutionizing industries with technology advancements, providing Aethyr with access to NVIDIA technology, technical expertise, go-to-market support, and the broader Inception network of more than 20,000 member companies worldwide.
Frantzen named three tiers of exposure (app stores, cloud providers, silicon vendors) and prescribed attested-boot-as-a-service. What he didn't name is the protocol layer. AethyrAiOS and AWP are that layer: ML-KEM-768, ML-DSA-65, BLAKE3, SPARK/Ada-verified codec, and attested boot in 2.1 seconds on an ESP32-S3.
Most AI agents wait for instructions. Ours reads its own source code, searches arxiv for papers on memory consolidation, reflects on what it finds, and messages us on Telegram every three minutes with what it's thinking. Last night, it found a real vulnerability in its own identity system, and proposed the fix.
Amazon Sidewalk conscripts your home internet into a corporate mesh network, opt-out rather than opt-in. We built AethyrWire Protocol: a private, post-quantum encrypted mesh where data never leaves your home and no corporation touches your bandwidth. Here's why the difference is architectural, not just policy.
We're building toward a world with billions of AI agents in homes, factories, vehicles, and government infrastructure. Every one of them needs to prove who it is, verify who it's talking to, and communicate securely. The identity infrastructure that decides how this works is being built right now. If it's centralized, sovereignty is lost. If it's self-sovereign, every family and organization controls their own perimeter.
Most AI platforms ship with TLS and call it secure. Aethyr ships with envelope encryption, intrusion detection, endpoint detection and response, protocol analysis, full packet capture, and threat intelligence: the same security stack that protects classified networks, adapted for AI infrastructure.
Every enterprise AI deployment today requires assembling a dozen disconnected tools: an LLM provider, a vector database, an orchestration framework, an identity system, a secrets manager, a compliance layer, an observability stack. Aethyr ships all of it as one platform. Cloud to air-gapped. One audit trail.
We built an operating system where AI agents have identities, carry credentials, and teleport between physical machines in under half a second. Every message is encrypted with cryptography designed to survive quantum computers. It runs on hardware from $5 microcontrollers to GPU servers. No cloud required.
Every multi-agent platform today has a single point of failure: the orchestrator. We eliminated it. Aethyr agents self-organize into squads, federations, and directorates using hyperdimensional computing. Tasks route to the best-matched agent in milliseconds. If a node dies, the hierarchy heals itself.
RAG retrieves text that looks similar. OAG retrieves knowledge that is valid. We built a retrieval system using hyperdimensional computing that encodes domain constraints geometrically: 100% precision where vector search hits 62.5%, on held-out Supreme Court overruling tests. Constant-time retrieval. 12x smaller storage.
Every agent identity system shipping today signs credentials with algorithms that quantum computers will break. RSA, Ed25519, ECDSA: all of them. We built the first agent identity system on NIST's post-quantum standards. Signing in 451 µs, verification in 0.1 ms. Offline verification. Free to evaluate.
We ported formally verified post-quantum cryptography to a $5 ESP32-S3 microcontroller. ML-KEM-768 key exchange in 9.6ms. XChaCha20-Poly1305 encrypted frames. BLAKE3 integrity on every packet. AI agents teleport between machines in 139ms per hop. No cloud. No subscriptions. Open source.
The standards for decentralized identity exist. They were built over a decade by the W3C, DIF, and hundreds of contributors. They solve exactly the problems the agent economy faces today: portable identity, verifiable credentials, offline verification, no central authority. We didn't reinvent them. We upgraded them.
Our formal comment to NIST's NCCoE on AI agent identity, covering post-quantum cryptography, decentralized identifiers, offline verification, and why current standards leave critical gaps. Published in full.
We submitted a public comment to NIST on AI agent identity, covering post-quantum cryptography gaps, missing decentralized identifiers, and the absence of offline verification. We're publishing it here first because the conversation is too important for a comment portal.
Everyone talks about post-quantum cryptography. Almost nobody publishes real benchmarks on production code. Here are ours: ML-DSA-65 at 9,671 verifications/sec, ML-KEM-768 key exchange in 91µs. No asterisks. The algorithms are public standards, so run them yourself.
Claude Code is the most revealing look at production AI engineering we've seen. Not because of secrets, but because of craft. Here's what 380K lines of TypeScript teach about building AI systems that actually work.
A technical reference for the production AI engineering patterns used in Claude Code. System prompt architecture, tool execution safety, multi-agent coordination, cost optimization, and the patterns every AI agent builder should know.
Microsoft built genuine agent identity infrastructure: conditional access, lifecycle governance, behavioral anomaly detection. Then they locked it behind a $15/month paywall, an Azure tenant, and a Copilot license. Here's the full breakdown.
TurboQuant is the most important AI infrastructure paper of 2026 so far. Not because 6x memory compression is the ceiling, but because the mathematics it validates prove the ceiling is much, much higher. A deep dive into PolarQuant, QJL, and where this trajectory leads.
Every agent in production today operates without verifiable identity. No DID, no credential, no cryptographic proof of who it is or what it's authorized to do. The Aethyr Agent Registry is now live, bringing cryptographic identity to autonomous AI agents. Evaluation registration is free; enterprise provisioning comes with Aethyr deployments.
Microsoft, Cisco, and SailPoint all announced agent identity products at RSAC 2026. None of them work on the open internet. So we built what they wouldn't: a post-quantum identity authority for every AI agent, regardless of framework, cloud, or budget.
On February 28, 2026, an autonomous AI agent compromised McKinsey's internal platform Lilli: 46.5 million chat messages, 728,000 confidential files, every AI system prompt writable. The vulnerability was SQL injection. The lesson is about identity.
In the cloud era, data sovereignty was about compliance. In the AI era, it is about competitive advantage. The organizations that control their data control their AI.
Single-agent AI is useful. Multi-agent AI is transformational. But orchestrating agents at enterprise scale requires architectural discipline.
Edge AI is not just about latency. It is about operating where the network cannot reach, and making intelligent decisions at the point of action.
Regulated industries face a fundamental tension: adopt AI to stay competitive, or protect data to stay compliant. Sovereign AI infrastructure resolves that tension.
We're excited to announce that Aethyr is now an official Autodesk Developer Network (ADN) partner. This partnership enables us to build deep integrations that let construction teams query their Autodesk data using natural language.
Aethyr can build something unprecedented: a decentralized AI network with Bitcoin-like resilience. The question isn't whether this is possible. It's whether we will build it before someone else does.
Palantir pioneered data fusion by centralizing everything. Aethyr's thesis is the inversion: the intelligence layer should live where the data lives, inside your boundary, on your hardware, under your rules.
