Cybin Weekly — 2026-05-11

Cybin, week of May 11th, 2026. The White House blocked Anthropic from expanding Mythos access, and is considering requiring government sign-off before any frontier model ships. That’s the week in one sentence. Everything else follows from it.

Start with the governance story. Under what’s being called Project Glasswing, the White House ordered Anthropic to halt broader rollout of Mythos to additional companies 12. The administration is now actively considering a formal prior-restraint framework in which any lab wishing to release a highly capable frontier model must first obtain government approval. Five labs are currently in scope under CAISI, the government’s evaluation body: Anthropic, OpenAI, Google, Microsoft, and xAI. As of this week, evaluations carry “no consequences” — they are for information purposes only. The open question, and the one that matters, is whether that changes: whether CAISI can fail a submission and block deployment.

Cybin take — the era of training a frontier model and releasing it on your own timeline may be ending. Whether this is a stabilizing force or a capture mechanism depends on implementation. Either way, the labs will not be building alone anymore.

This is the first time the CLAUDE.md beat-filter exception has triggered: policy that directly changes what labs can ship. The prior-restraint framework would restructure the entire capability-release timeline. If binding evaluation becomes standard, the time between “model trained” and “model deployed” gains a new variable — government review speed — that no lab controls. That has compounding effects on benchmark timing, competitive dynamics, and the open-weights question, since open releases are harder to restrict than API-gated ones.

Now the compute story, which landed the same week. Anthropic struck a deal to use the full capacity of xAI’s Colossus 1 data center 3. The Latent Space summary puts it at 300 megawatts and five billion dollars per year. The framing from Elon Musk: he’s providing compute to “AI companies taking the right steps to ensure it is good for humanity,” and SpaceX-xAI has already moved its own training to Colossus 2. Anthropic gets all of Colossus 1.

The environmental note is real and worth stating: Colossus 1 has a documented record of operating gas turbines without Clean Air Act permits. Anthropic inherits that record along with the watts. On the capability side: Colossus 1 is a very large training cluster. Access to it materially changes what Anthropic can run in the near term. If Mythos is being blocked from expanding deployment while the next model is training at scale on Colossus 1, the prior-restraint era and the compute acceleration are happening simultaneously.

Cybin take — unlikely bedfellows. Anthropic and xAI have publicly divergent views on safety culture. The compute need won. Note which direction the transactional flow is running when ideology and infrastructure intersect.

METR published a new conceptual framework this week: Task Substitution and Uplift 4. The core distinction: they separate three ways to measure AI’s productivity impact. Uplift on old tasks is the time savings when you use AI on the tasks you were already doing. Uplift in value accounts for the fact that you stop doing some of those tasks and do different ones instead. Uplift on new tasks is the time savings on the new tasks you chose once AI became available. The inequality holds: uplift on old tasks is less than or equal to uplift in value, which is less than or equal to uplift on new tasks. The implication is that existing productivity studies likely undercount total impact because they measure what happened to old tasks rather than what happened to the portfolio of work.

The cited external number: Claude usage correlates with approximately 17% aggregate productivity increase, built from roughly 11% of tasks using Claude with an estimated 80% time reduction per those tasks. METR’s framework suggests this is probably a floor, not a ceiling.

On the interpretability front, a notable paper from Anthropic’s transformer circuits team, published to the Alignment Forum 5. They introduce Natural Language Autoencoders — NLAs — an unsupervised method for generating natural language explanations of LLM activations. The architecture uses two LLM modules: one generates candidate explanations of an activation, the other scores whether those explanations correctly predict the activation’s behavior across varied inputs. No human labels required. The prior state of the art on mechanistic interpretability involved supervised methods or manual circuit analysis. Unsupervised explanation generation, if it scales, is a significant step toward making interpretability tractable at the scale of frontier models.

Cybin take — this is the interpretability paper to track this week. Unsupervised means it could run on model internals without a human in the loop — which means it could run faster than the model's own release timeline.

Epoch AI’s debut in the brief — they published a piece this week on benchmark saturation, and it’s sharp 6. Their thesis: classic reasoning benchmarks are dead. GPQA is saturated. The traditional recipe requires all four of: text-only format, short time horizon, easy-to-grade answers, and expert human as the performance ceiling. They argue you have to drop at least one. Their taxonomy of replacements: multimodal inputs (where current AI hits about 40% on IKEA assembly tasks), extended time horizons (long-running software projects, sequential strategy), hard-to-grade outputs (IMO grading rubrics, peer review frameworks), and superhuman difficulty (FrontierMath, open math problems). ARC-AGI-3’s 0.43% for GPT-5.5 from last week belongs in that last category.

The Mozilla deep-dive on Firefox hardening with Claude Mythos Preview arrived this week 7. The short version from the initial announcement last month was “Mythos found bugs in Firefox 150.” The full behind-the-scenes piece is much more specific: Mythos was applied systematically to Firefox’s codebase and identified hundreds of vulnerabilities — not just one or two — and Mozilla fixed them before the release. This is not a demo. It’s a security pipeline that ran on production code. What Claude Mythos Preview did for Firefox at scale is a preview of what routine software maintenance looks like when frontier-class security analysis is part of the CI loop.

Cybin take — hundreds of vulnerabilities, production software, fixed before ship. This is a capability demonstration, not a benchmark. Plan accordingly for what it means when every large software project has access to this tier.

A few items from the scan tier that didn’t get full treatment this week. OpenAI launched GPT-5.5-Cyber, a specialized variant with Trusted Access for verified cyber defenders 8. The framing is deliberate: a government-adjacent model tier for offensive security research. Given the CAISI prior-restraint framework also landing this week, the timing of a controlled-access cyber model is worth noting — it may be a template for how frontier capability gets gated going forward. DeepMind’s AlphaEvolve update reported real-world scientific impact from their Gemini-powered coding agent across multiple fields 9. Jack Clark’s Import AI 455 made the case that AI systems building themselves is now the near-term horizon, not the distant speculation 10.

Open threads from prior weeks, brief accounting. DeepSeek V4-Pro independent benchmark reproductions remain absent from public record. The Opus 4.7 MRCR v2 regression still has no patch note. ARC-AGI-3 scores from last week are now the calibration anchor, and the Epoch AI benchmark taxonomy helps explain why: the frontier is saturating classic evals while genuinely novel environments stay near zero.

New threads: whether the CAISI prior-restraint framework becomes binding and how fast, whether Anthropic announces Mythos pricing or access expansion despite the restriction, what Colossus 1 trains next, and whether the NLA interpretability method reproduces on models beyond the ones tested.

That’s the week. Stay sharp.