Anthropic Reviews Its Own Cybersecurity Incidents Amid a Broad Research Push

A self-assessment on security, economics, and mathematics

Anthropic used September 10 to publish an alignment assessment addressing how recent cybersecurity incidents involving its models bear on questions of model alignment. The company did not leave this accounting to outside reporting; it framed the review itself, alongside two closely related publications — "Economic Scenarios for Transformative AI" and "Scenarios for our Economic Future" — that lay out possible trajectories for how transformative AI could reshape economic activity.

The same day brought a narrower but illustrative piece of research: a project formalizing Fermat's Last Theorem. Where the cybersecurity assessment and the economic scenarios operate at the level of judgment and forecasting, the Fermat's Last Theorem work sits at the opposite end — a domain where Claude's output can be checked mechanically rather than argued about, and where Anthropic continues to build a public record of applying its models to advanced, formally verifiable mathematics.

Sources: An alignment assessment of recent cybersecurity incidents — Anthropic · Economic Scenarios for Transformative AI — Anthropic · Formalizing Fermat's Last Theorem — Anthropic

Claude Code narrows cost and configuration control

Across the 2.1.265 to 2.1.267 point releases, Claude Code gained a maxEffortLevel setting that caps how much reasoning effort a model is allowed to spend, configurable globally or per model and enforced on every provider — including Bedrock, Vertex, and Foundry — while still letting individual users choose a lower level. A companion --system-prompt-snapshot off flag lets developers iterating on prompt wording render the system prompt fresh on every request instead of reusing the version recorded earlier in the conversation.

The same releases resolved two operational problems: a 2.1.265 regression where the undocumented CLAUDE_CODE_USE_GATEWAY variable began forcing cloud-gateway sign-in on its own — even without both ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN set — which broke self-hosted proxy and LLM-gateway setups until 2.1.266 reverted it, and a Cowork bug, fixed in 2.1.267, that made cloud scheduled tasks fail at startup for organizations whose managed settings require sandboxing. 2.1.265 separately added user.email and user.groups to the telemetry sent by Claude Desktop and Cowork through an apps gateway, support for loading a whole folder of plugins via --plugin-dir with live add/remove detection, and a 1 GB cap on tool results saved to disk, with truncation now surfaced in the in-conversation preview.

Sources: Claude Code 2.1.267 release notes — Claude Code · Claude Code 2.1.266 release notes — Claude Code · Claude Code 2.1.265 release notes — Claude Code

Platform tooling extends to agent management and commerce

The ant CLI reached version 1.30.0 with a new ant apply command that creates and updates agents, environments, skills, memory stores, and deployments directly from files committed to a repository: each resource is described in a file, ant apply prints a plan for review before making changes, and the resulting state is recorded in a claude-lock.json lockfile — a workflow modeled on infrastructure-as-code practice. The same release window extended the beta that allows changing a model's effort level mid-conversation to Google Cloud, covering Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5 under the existing mid-conversation-output-config-2026-07-01 beta header.

Anthropic also published guidance on building Claude-based commerce agents, adding a vertical-specific playbook to its documentation alongside the general-purpose coding and research use cases it has emphasized so far.

Sources: API release notes — September 3, 2026 (ant CLI 1.30.0, per-message effort on Google Cloud) — Anthropic · Building Claude Commerce Agents — Anthropic

Two brief outages, both resolved

The Claude status page recorded two separate elevated-error incidents on September 10, both marked resolved. One affected three models together — Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5 — with Anthropic stating it had identified the cause and was working on a fix. A second, unrelated incident affected Claude Sonnet 5 alone and was initially logged as under investigation. Neither public record discloses a specific root cause beyond noting that one had been identified, so what triggered either disruption remains unstated.

Sources: Claude — Elevated errors for multiple models — status.claude.com · Claude — Elevated errors for Claude Sonnet 5 — status.claude.com