"Claude autonomously broke post-quantum encryption experts missed for 2 years" — AlphaSignal

Why this is in the vault

Dated marker for a concrete instance of Claude (Mythos) autonomously producing a novel cryptanalytic result — a capability-frontier data point for the ongoing AI-capability watch, not a vendor claim to take at face value.

The core argument

Two results, both framed as Anthropic-reported: Claude Mythos Preview found a hidden mathematical shortcut in the HAWK post-quantum signature scheme in 60 hours — a scheme that had survived two years of expert human review — cutting its effective security strength in half. Separately, it sped up attacks on a weaker research variant of AES by 200-800x via a technique it named "Mobius Bridge," working autonomously with minimal human input. Each result reportedly cost roughly $100,000 in API usage. AlphaSignal's own caveat, stated directly in the body: neither breaks anything in production — HAWK isn't deployed yet, and the AES attack only hits a weaker research variant, not the real standard. The framing is "AI as a serious tool for stress-testing encryption before deployment," not "Claude broke the internet."

Curation section

Top news (full write-up): MCP's biggest update yet goes stateless — persistent sessions dropped, any server instance can now handle any request, unlocking serverless/load-balanced deployment (Cloudflare Workers named explicitly), plus an Extensions framework (MCP Apps, Tasks), Enterprise Managed Auth, and a formal deprecation policy. Updated SDKs for Python, TypeScript, Go, C#.
Top news (full write-up): the HAWK/AES Mythos result above (Anthropic).
Top news (full write-up, sponsored — Datadog): Cursor launches a ₹649/month ("Cursor Start") India-specific plan with Grok 4.5 and cloud agents, excluding frontier OpenAI/Anthropic models and Pro-only features.
Signals (6 one-liners): Claude built a full AAA-style space game in one 24-hour coding session · TimescaleDB/Tiger Data extends Postgres for millisecond queries (sponsored placement) · Microsoft open-sourced a graph-based RAG tool (34,958 stars) · a study found AI-written books now take ~20% of Amazon sales · LiquidAI released 230M/350M multilingual encoder models · an open-source text-to-CAD tool hit 10,000 stars.

Two deep-fetches were plausible candidates (the Mythos/encryption story, the MCP-stateless story) but both were skipped: the newsletter's own body already contains the load-bearing facts and caveats for each, and neither links to a primary Anthropic/MCP-spec post with material beyond what's quoted above.

Mapping against Ray Data Co

Direct hit on the AI-capability/safety watch this agent runs continuously on Ben's behalf while pursuing the Anthropic Claude Certified Architect cert via phData (target 2026-11-22): this is a specific, dated instance of a Claude model doing autonomous offensive security research (cryptanalysis) with "minimal human input," which is exactly the shape of capability the RSP/Glasswing threshold conversations (see Related) are about — and it lands one day after the MCP-stateless update, which is itself a capability-adjacent infrastructure shift (agents that can run detached from a session, scale horizontally, live behind standard load balancers). Both threads matter for how this agent harness itself is built: an autonomous Claude-based agent running detached research loops (deep-research, family-research-round, investing-edgar-watch) is a smaller-scale version of the same "give the model a hard problem and minimal supervision" pattern AlphaSignal is reporting at $100K/result scale. Worth a light touch-base at the next cert-study or capability-watch pass, not an action item today — the newsletter's own caveat (research-variant only, nothing in production broken) keeps this from crossing into a decision-worthy threshold.

[[2026-04-08-innermost-loop-mythos-glasswing]] — the original Mythos capability-threshold marker (Project Glasswing coalition, RSP discourse, "post-quantum" tag); this note is the follow-on autonomous-cryptanalysis data point for that same model line
[[2026-04-13-stratechery-mythos-muse-compute]] — Ben Thompson's strategic read on the same Mythos launch, useful cross-check against AlphaSignal's more breathless framing