Tokens & Signals · Monday, June 22, 2026

Mythos Breach: Did Anthropic Just Break Security?

gb300claudecodex-securitymythosgb200gpt-5.5-codexclaude-opus-4.8-xhigh-effortgemini-3.1-proglm-5.2fugu-ultraspacexreflection-ainvidiaanthropicpersonaopenaisk-hynixzhipu-aisakana-aimetacompute-infrastructureidentity-verificationsecurity-vulnerabilitiesexport-controlsmemory-marketagentic-scalemodel-benchmarkingmulti-agent-orchestrationinference-optimizationkarpathysamaemaborotkunal-shahbernie-sanders
Tokens & Signals for 6/22/2026. We scanned ~1,200 Twitter accounts (1411 tweets), 13 subreddits (88 posts), Hacker News (11 stories), 7 newsletter posts, 5 podcast episodes, 207 Discord messages, and leaderboard data for you. Estimated reading time saved: ~14 hours.

TLDR & AI Twitter Recap

* SpaceX locks in massive compute: A $6.3B deal with Reflection AI ($150M/month through 2029) brings Nvidia GB300 chips to the Colossus II center. x.com/ns123abc/status/2069090675451142407(https...)

* Anthropic's ID shift: Starting July 8, 2026, consumer Claude users will need to hand over a government ID and a biometric selfie via Persona. x.com/lennysan/status/2069138871779238064(https...)

* OpenAI's auto-patching: The new Codex Security plugin doesn't just flag vulnerabilities — it actually fixes them. That's a meaningful leap. x.com/sama/status/2069121360744550796(https://x...)

* Mythos model's NSA breach: Reports allege Anthropic's model cracked classified NSA systems in hours, and that performance apparently spooked the U.S. enough to trigger the recent export controls. x.com/AndrewCurran_/status/2068748019030483365(...)

* Memory market crunch: Prices are spiking 40-50% QoQ as hyperscalers hoard HBM; SK hynix has now overtaken Samsung in market cap. x.com/jukan05/status/2068895341706109191(https:...)

* @karpathy on the DeepSeek WoW bots: "The dead internet theory just became a product roadmap." 1,800 AI bots running around a game server, trading and fighting — that's agentic scale in the wild. reddit.com/r/ChatGPT/comments/1ubl0gc/the_dead_...)

* Software-led speed: Simple software optimizations on GB200 clusters are pushing throughput up 2.5x, which is a good reminder that the best GPU is the one you're already using more efficiently. x.com/SemiAnalysis_/status/2069103359785590969(...)

* @emaborot on the Mythos incident: "If true, this is the scariest AI story of the decade. Hours, not weeks. That's not a red-team — that's a demonstration." x.com/AndrewCurran_/status/2068748019030483365(...)


Go deeper on what matters to you

Tap to expand

Best to Build With Today

* Codinggpt-5.5-codex is currently the top performer on LiveBench coding tasks.

* Reasoningclaude-opus-4.8-xhigh-effort is the strongest reasoning model available right now.

* Chatgemini-3.1-pro holds the top ELO spot on the Chatbot Arena for general chat.

* Open-sourceGLM-5.2 is the undisputed leader for open-weights, especially for agentic coding.


Deeper Dives

💼 Industry & Business

SpaceX Signs $6.3B Compute Deal

SpaceX is going all-in on infrastructure, leasing capacity to Reflection AI ($25B valuation) in a $6.3B deal for Nvidia GB300 chips at the Colossus 2 data center in Memphis through 2029. 📱 Twitter

Anthropic Identity Verification

Starting July 8, 2026, Anthropic is requiring consumer Claude users to verify their identity with a government ID and a biometric selfie through Persona Identities. Enterprise users are off the hook. 📱 Twitter · 💬 Reddit

Memory Market Price Surge

AI demand is creating a brutal supply crunch — DRAM and SSD prices are expected to climb 40-50% in Q3 2026. Memory now accounts for 23% of device costs, so if you were planning a hardware upgrade, you might want to do it soon. 📱 Twitter

Mythos Model and NSA Breach

Reports claim Anthropic's Mythos model punched through classified NSA systems in just hours during an authorized red-team test. That performance apparently became a primary catalyst for the June 12 U.S. export controls. 📱 Twitter · 💬 Reddit

Bernie Sanders' $7T AI Plan

Senator Sanders unveiled a sweeping policy framework aimed at putting more public control over the AI industry. Whether you agree with it or not, AI is clearly a top-tier political issue now. 💬 Reddit

🧠 Models & Research

GLM-5.2 Emerges

Zhipu AI's 753B parameter MoE model comes with a 1M-token context window and a novel "index share" mechanism that cuts per-token compute costs significantly. It's the new benchmark to beat for open-weights. 📱 Twitter · 💬 Reddit

Inference Optimization via Software

Turns out you don't always need more GPUs — recent deployments show that software tweaks like quantization and KV cache management can boost token throughput by 2.5x on their own. Software is quietly becoming the biggest lever in AI cost-efficiency. 📱 Twitter

🚀 Products & Launches

OpenAI Codex Security

OpenAI's new Daybreak plugin goes well beyond flagging issues — it scans entire codebases, finds vulnerabilities, then automatically generates, verifies, and deploys the patches. That last step is the interesting one. 📱 Twitter · 🔶 Hacker News

Sakana AI's Fugu Ultra

Fugu Ultra is a multi-agent orchestration layer that routes tasks through a swappable pool of specialized agents, all through a single API. Smart way to avoid getting locked into any one vendor. 📱 Twitter


Funding & Deals

* Reflection AI locked in a $6.3B compute lease from SpaceX to access GB300 chips through 2029.

* Meta invested $1 billion in CRED to bring founder Kunal Shah in to lead WhatsApp's global product strategy.

* Google DeepMind formed a $75 million research partnership with film studio A24 to build creative AI tools.


Launches

* Codex Security — OpenAI's full vulnerability remediation suite for enterprise codebases.

* Fugu Ultra — Sakana AI's multi-agent API for frontier-model orchestration.


Closing thought: Between the Mythos NSA breach and 1,800 autonomous bots running an entire game server, the line between "research toy" and "autonomous agent" has officially disappeared. The policy response is coming — and it's going to be heavy.