AI progress — 21 October 2026

The most consequential item today is outside the model race: eight eMudhra intermediate CA private keys have been compromised, and they chain to roots trusted by the Mozilla, Microsoft, Apple and Chrome programs. Anything that bundles its own trust store needs an update. On the AI side, Anthropic is reported to be pitching a safety bill and to be short of cloud compute ahead of a planned November IPO, and Meta paused chat sharing in Muse after shared links turned up in public DNS records.

New models and papers

• Sage-A1.5 — a 35B Mixture-of-Experts agentic model; the report says it learns to organize work and revise its approach during execution. Why it matters: these are the authors’ claims from a technical report, so treat the long-horizon results as unverified until independently tested.
• ALTK-Evolve — IBM Research describes a learned-memory approach that it says tops the τ²-bench customer-service leaderboard without a frontier model. Why it matters: if the claim holds, agent quality for support workloads may come from memory rather than a bigger model, which changes the cost of running them.
• Focused Omni-Modal GUI Agents — a paper arguing GUI agents should not expose the full screenshot, audio, video and history context at every step, since relevance is sparse and shifts. Why it matters: it targets context clutter in computer-use agents; the abstract is what we have here, not the results.

New tools and software

• GitHub package firewall (public preview) — a GitHub control aimed at malware that runs at package install time and harvests credentials on CI/CD hosts and developer machines. Why it matters: install-time attacks hit CI runners before any of your code runs, so this is a new place to block them if you use GitHub-hosted workflows.
• Claude Code v2.1.305 — adds sandbox.allowNestedUserNamespaces for Linux sandboxes (Chromium, rootless podman) and a usage-limit recheck setting for unattended retry mode. Why it matters: the release notes themselves warn that the namespace option lets commands get around the sandbox’s file protections, so enabling it is a deliberate trade-off.
• eMudhra intermediate CA compromise — private keys of eight eMudhra intermediate CA certificates are compromised, per the post Simon Willison highlights. Why it matters: any software that bundles trusted certificates needs updating, including agent and CI images with their own CA bundles.

Overall digest

• Anthropic’s proposed AI safety bill — TechCrunch reports a proposal that could block a model if 50+ people could be killed, while barring federal agencies from stopping a release without a judge’s permission. Why it matters: it is TechCrunch’s reading of a pitch, not law; the judge-approval clause is the part to watch for anyone shipping frontier models.
• Meta pauses Muse chat sharing — researchers at Claroc reportedly found more than 6,800 shared Muse links through public DNS records, and Meta paused the feature. Why it matters: shared-conversation URLs that are discoverable are effectively public; check what your own assistant’s share links expose.
• Anthropic’s cloud partnerships — reportedly struggling to secure enough compute from cloud providers that see it as a competitor, with a November IPO planned. Why it matters: this is a single report; if accurate, it bears on capacity for anyone building on Anthropic’s API.